SEMINAR

Speaker: Anindita Maiti (Perimeter Institute)

Title – Neural Network Field Theory and Wilsonian RG: Bridging Physics and Machine Learning

Abstract – Machine learning (ML) models have become indispensable tools across diverse domains, including theoretical physics. However, their utility hinges on trustworthiness—precision, interpretability, and robustness—especially in applications to complex physical systems. In this talk, I will present two foundational contributions that aim to bridge physics and machine learning by leveraging concepts from quantum field theory (QFT) and renormalization group (RG) flows. First, I will introduce the Neural Network Field Theory (NNFT) correspondence, which maps neural network (NN) architectures, the backbones of ML, to QFTs. This correspondence provides a framework to embed symmetries and interactions of QFT actions directly into NN architectures at initialization, improving NN interpretability and robustness. By drawing analogies with field interactions, I will show how violations of the Central Limit Theorem (CLT) in NN parameter distributions can be harnessed to design more expressive and physically meaningful models. Second, I will discuss a Wilsonian RG framework for supervised learning performance, which tracks the evolution of noise in NN predictions as data features are coarse-grained. This approach offers a novel perspective on how precision emerges or degrades based on data feature scales, providing theoretical tools to identify regimes of high trustworthiness in ML outputs. By integrating these insights, this talk demonstrates how theoretical physics not only inspires but also fundamentally reshapes the design and understanding of machine learning models.