![\"Kevin](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/08\/Zhang_Headshot.jpeg\")
<\/a>Zhengkang (Kevin) Zhang, assistant professor, Department of Physics & Astronomy<\/p>\n\u201cPeople used to say machine learning is a black box\u2014you input a lot of data and at some point, it reasons and speaks and makes decisions like humans do. It feels like magic because we don\u2019t really know how it works,\u201d said Zhang. \u201cNow that we\u2019re using AI across many critical sectors of society, we have to understand what our machine learning models are really doing\u2014why something works or why something doesn\u2019t work.\u201d<\/p>\n
As a theoretical particle physicist, Zhang explains the world around him by understanding how the smallest, most fundamental components of matter behave in an infinitesimal world. Over the past few years, he\u2019s applied the tools of his field to better understand machine learning\u2019s massively complex models.<\/p>\n
Scaling up while scaling down costs<\/p>\n
The traditional way to program a computer is with detailed instructions for completing a task. Say you wanted software that can spot irregularities on a CT scan. A programmer would have to write step-by-step protocols for countless potential scenarios.<\/p>\n
Instead, a machine learning model trains itself. A human programmer supplies relevant data\u2014text, numbers, photos, transactions, medical images\u2014and lets the model find patterns or make predictions on its own.<\/p>\n
Throughout the process, a human can tweak the parameters to get more accurate results without knowing how the model uses the data input to deliver the output.<\/p>\n
Machine learning is energy intensive and wildly expensive. To maximize profits, industry trains models on smaller datasets before scaling them up to real-world scenarios with much larger volumes of data.<\/p>\n
\u201cWe want to be able to predict how much better the model will do at scale. If you double the size of the model or double the size of the dataset, does the model become two times better? Four times better?\u201d said Zhang.<\/p>\n
A physicists\u2019 toolbox<\/p>\n
A machine learning model looks simple: Input data\u2014>black box of computing\u2014>output that\u2019s a function of the input.<\/p>\n
The black box contains a neural network, which is a suite of simple operations connected in a web to approximate complicated functions. To optimize the network\u2019s performance, programmers have conventionally relied on trial and error, fine-tuning and re-training the network and racking up costs.<\/p>\n
\u201cBeing trained as a physicist, I would like to understand better what is really going on to avoid relying on trial and error,\u201d Zhang said. \u201cWhat are the properties of a machine learning model that give it the capability to learn to do things we wanted it to do?\u201d<\/p>\n
In a new paper<\/a> published in the journal Machine Learning: Science and Technology, Zhang solved a proposed model\u2019s scaling laws, which describe how the system will perform at larger and larger scales. It\u2019s not easy\u2014the calculations require adding up to an infinite number of terms.<\/p>\n Zhang applied a method that physicists use to track hundreds of thousands of terms, called Feynman diagrams. Richard Feynman invented the technique in the 1940s to deal with hopelessly complicated calculations of elementary particles in the quantum realm. Instead of writing down algebraic equations, Feynman drew simple diagrams\u2014every line and vertex in the diagram represents a value.<\/p>\n \u201cIt\u2019s so much easier for our brains to grasp, and also easier to keep track of what kind of terms enter your calculation,\u201d Zhang said.<\/p>\n<\/a>A portion of the Feynman diagrams used to solve a machine learning model.<\/p>\n