As modern life increasingly relies on machine learning, a form of artificial intelligence that learns from datasets without explicit programming, the technology remains poorly understood.
Zhengkang (Kevin) Zhang, an assistant professor at the University of Utah’s Department of Physics & Astronomy, is working to unveil the complexities of machine learning.
Machine learning has become ubiquitous, powering everything from self-driving cars to facial recognition systems, yet its mechanisms remain something of a mystery.
Zhang highlights the need for a deeper understanding of this technology, particularly as it becomes integral to various critical sectors in society.
He remarked, “People used to say machine learning is a black box—you 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’t really know how it works.”
Zhang approaches the challenge from his unique perspective as a theoretical particle physicist, whose work focuses on understanding matter at its most fundamental level.
With growing interest in machine learning, he has adopted tools from his field to analyze the highly complex models these systems employ.
Traditionally, computer programming followed a straightforward approach: a programmer specified detailed instructions to accomplish a task.
For instance, to create software capable of detecting irregularities on a CT scan, a programmer would construct intricate protocols for numerous potential scenarios.
Conversely, machine learning models operate differently. Instead of following instructions, they train themselves using data provided by a human programmer, including text, numbers, images, and transaction logs, allowing the model to independently identify patterns and make predictions.
Interestingly, while a human can adjust parameters to improve accuracy, the inner workings of the model—how it transforms input data into output—often remain opaque.
Additionally, machine learning is known for its high energy consumption and costs, prompting industries to first train models on smaller datasets before scaling them up for larger real-world applications.
Zhang poses an essential question: “We 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?”
Zhang’s tools as a physicist help illuminate these issues.
To explain how machine learning models work, one can visualize them as a straightforward process of inputting data into a ‘black box’ that eventually produces an output linked to that input.
The complexities hidden within this black box involve a neural network composed of interconnected operations designed to approximate complex functions.
Traditionally, programmers have relied on trial and error, which leads to significant costs when fine-tuning the network’s performance.
However, Zhang seeks a more systematic approach: “Being trained as a physicist, I would like to understand better what is really going on to avoid relying on trial and error. What are the properties of a machine learning model that give it the capability to learn to do things we wanted it to do?”
A recent paper Zhang published in the journal Machine Learning: Science and Technology tackles this issue by analyzing scaling laws of a proposed machine learning model.
These scaling laws indicate how the system will fare as it performs at larger scales, a task that is computationally challenging due to the need to sum an infinite number of terms.
To tackle this, Zhang employed a technique known as Feynman diagrams, originally developed by Richard Feynman in the 1940s for resolving complex calculations in quantum physics.
Instead of resorting to traditional algebraic equations, Feynman diagrams simplify the calculations into more digestible visual representations, where each line and vertex correspond to specific values.
Zhang utilized this method to extend the analysis of a model previously studied in research from 2022, deriving new and refined scaling laws that dictate the behavior of the model beyond the previously understood limits.
As society accelerates towards greater reliance on AI, a growing number of researchers are striving to ensure its responsible use.
Zhang firmly believes physicists can collaborate with engineers, computer scientists, and others to navigate the ethical implications of AI.
Notably, he cautions against the unintended consequences of technology: “We humans are building machines that are already controlling us—YouTube algorithms that recommend videos that suck each person into their own little corners and influence our behavior.”
Zhang concludes, emphasizing the critical nature of understanding these technologies: “That’s the danger of how AI is going to change humanity—it’s not about robots colonizing and enslaving humans. It’s that we humans build machines that we are struggling to understand, and our lives are already deeply influenced by these machines.”
image source from：attheu
