An Enquiry on similarities between Renormalization Group and Auto-Encoders using Transfer Learning

TL;DR

This paper explores the parallels between Renormalization Group and Auto-Encoders in neural networks, using Transfer Learning to compare coarse-graining with encoding-decoding processes, aiming to deepen understanding of learning mechanisms.

Contribution

It introduces a novel investigation of the relationship between RG and Auto-Encoders, employing Transfer Learning to analyze their similarities in neural network structures.

Findings

Identifies structural similarities between RG and Auto-Encoders.

Demonstrates how Transfer Learning can embed coarse-graining in neural networks.

Provides insights into the mathematical parallels underlying learning processes.

Abstract

Physicists have had a keen interest in the areas of Artificial Intelligence (AI) and Machine Learning (ML) for some time now, with a special inclination towards unravelling the mechanism at the core of the process of learning. In particular, exploring the underlying mathematical structure of a neural net (NN) is expected to not only help us in understanding the epistemological meaning of `Learning' but also has the potential to unravel the secrets behind the workings of the brain. Here, it is worthwhile to establish correspondences and draw parallels between methods developed in core areas of Physics and the techniques developed at the forefront of AI and ML. Although recent explorations indicating a mapping between the Renormalisation Group(RG) and Deep Learning(DL) have shown valuable insights, we intend to investigate the relationship between RG and Autoencoders(AE) in particular. We will use Transfer Learning(TL) to embed the procedure of coarse-graining in a NN and compare it with the underlying mechanism of encoding-decoding through a series of tests.