Network properties determine neural network performance.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Network Science

Background:

• Artificial neural networks (ANNs) are crucial in modern technology but lack systematic understanding.
• Current challenges in analyzing ANNs stem from complex configurations and data-dependent architectures.

Purpose of the Study:

• To develop a mathematical framework for analyzing ANN mechanisms.
• To introduce a universal metric for predicting model generalization capability.

Main Methods:

• Developed a framework mapping ANN performance to line graph network characteristics.
• Utilized edge dynamics of stochastic gradient descent differential equations.
• Derived a neural capacitance metric based on the mathematical framework.

Main Results:

• The neural capacitance metric universally captures generalization capability.
• Model performance can be predicted using only early training results.
• Demonstrated effectiveness on 17 ImageNet models across multiple datasets and a NAS benchmark.

Conclusions:

• Neural capacitance is a powerful indicator for model selection.
• The metric offers a more efficient approach compared to state-of-the-art methods.