Criticality Analysis of Artifical Neural Networks in Nuclear Physics

Machine learning methods, in particular deep learning methods such as artificial neural networks (ANNs), have become extremely useful tools in nuclear physics.

However ANNs often are treated as "black boxes", where their architecture (width, depth, and weight/bias initialization) and training algorithm is decided empirically based on what allows them to learn best. To actually investigate the principles of neural network architecture in an non-empirical, method based way, some have turned to performing criticality arguments with renormalization group flows in terms of the hyperparameters for weight/bias initialization, training rates, and the ratio of depth to width. These criticality arguments are meant to tune these hyperparameters to give an effective theory of the neural network that allows for the study of properties of neural networks so they can have a more ideal architecture chosen from the beginning, giving methodology to improve many projects in nuclear physics utilizing ANNs as tools.

In this talk, we demonstrate the advantages of the criticality analysis for some basic nuclear physics examples.