A novel mapping between machine learning and phase transition

Phase transition is an important physical concept which represents the transformation from one thermodynamic state to another. It has a wide application in physics such as condensed matter physics, particle physics and astrophysics. Machine learning is the study of computer algorithms that improve automatically through experience. It is also a very active research field with the success of deep learning in recent years. This paper visually gives the evidence that the training process of supervised learning is quite similar to phase transition. Accordingly, it is possible to identify the central concepts of phase transition theory, such as symmetry breaking or critical point in machine learning. The Ising model and an autoencoder of the MNIST dataset are used as two examples to explain the “phase transition” in the training process of machine learning. This novel mapping between machine learning and phase transition brings a new method to understand machine learning and makes it possible to solve machine learning problems from a physical perspective.