Five fields that usually ignore each other turn out to be the same math.

Scientists in different fields often study the same ideas without realizing it because they use different words. Over the last decade, researchers created powerful methods to learn complex patterns from massive data. This review shows how five separate areas actually share a single core thread: control theorycontrol theoryThe study of how to make a system behave a desired way by feeding it inputs over time., optimal transportoptimal transportThe math of moving resources from one place to another as cheaply as possible., probabilistic inferenceprobabilistic inferenceThe process of figuring out hidden causes from observed data, using probabilities., non-equilibrium thermodynamicsnon-equilibrium thermodynamicsThe physics of systems that are not in a stable, balanced state, like heat flowing through a rod., and machine learning. At the heart of these fields lies one shared method: optimizing free-energy-like mathematical formulas while following rules about physics or statistics. Starting with basic physics ideas that require no prior background, this paper connects these subjects. It shows how this unified viewpoint works in real machine learning tasks, specifically within reinforcement learning, variational inference, and generative modeling.

Why it matters

Anyone who builds learning systems, controls robots, or studies how heat and information flow can borrow ideas from the other fields once the shared structure is visible. This review makes that shared structure explicit, so a trick from thermodynamics can inform a machine-learning algorithm, and vice versa.

Questions answered