Toward Rigorous Parameterization of Underconstrained Neural Network Models Through Interactive Visualization and...

Area of Science:

• Computational Neuroscience
• Systems Neuroscience

Background:

• Neural network models often face challenges with non-unique solutions and multiple trajectories to final states.
• Parameter fitting in these models can lead to conclusions limited to local minima without systematic exploration.
• Neuroscience simulations, like those in other fields, require robust methods for handling complex parameter spaces.

Purpose of the Study:

• To develop an interactive tool for visualizing and steering parameters in neural network simulation models.
• To specifically address the complex parameter search scenario of connectivity generation in neural networks.
• To enable researchers to explore parameter spaces more effectively and gain a deeper understanding of neural network models.

Main Methods:

• Development of an interactive tool for real-time visualization and parameter steering during neural network simulations.
• Focus on connectivity generation as a key use case for parameter exploration.
• Incorporation of user-guided steering to achieve targeted network properties, such as specific mean activity levels.

Main Results:

• The tool allows researchers to interactively steer connectivity generation parameters to create networks with targeted mean activity.
• It facilitates exploration of various connectivity and neuron variables beyond the presented use cases.
• Demonstrated reduction in model assessment turnaround times through interactive visualization during simulation.

Conclusions:

• The developed tool enables interactive exploration of parameter spaces, improving the understanding of neural network models.
• It provides a method to grapple with the problem of non-unique network solutions and trajectories.
• The interactive approach enhances efficiency in assessing and developing neural network models.