Understanding the Mysteries of Machine Learning Through Physics
From self-driving cars to facial recognition, modern life is increasingly reliant on machine learning, a form of artificial intelligence (AI) that learns from datasets without explicit coding. The technology permeates various facets of society, yet our understanding of its underlying mechanisms is still in its infancy.
The Black Box of Machine Learning
Zhengkang (Kevin) Zhang, an assistant professor in the University of Utah’s Department of Physics & Astronomy, emphasizes that while machine learning can seem magical, it often feels like a “black box.” “You input a lot of data, and at some point, it reasons and makes decisions similar to humans,” Zhang explains. “But we need to understand why something works and why it doesn’t.”
Zhang has uniquely positioned himself to shed light on this “black box” by applying principles from theoretical particle physics to machine learning models. His work demonstrates how physicists can unravel the complexities and enhance our understanding of AI.
Traditional Programming vs. Machine Learning
Traditionally, programming a computer requires detailed, step-by-step instructions. For instance, creating software to spot irregularities in CT scans involves programming countless scenarios. In contrast, machine learning models train themselves by analyzing large datasets—text, photos, medical images—allowing them to identify patterns independently.
A human programmer can adjust parameters to improve accuracy, but the inner workings of the model often remain opaque. This opacity can lead to significant costs; machine learning can be energy-intensive and expensive, compelling industries to train models on smaller datasets first before escalating to larger, real-world applications.
Predicting Performance at Scale
Zhang points out the challenge: “We need to predict how much better the model will perform as it scales. If you double the model size or the dataset, does the model become twice as good? Four times better?” Understanding these scaling laws is essential for maximizing efficiency and effectiveness in machine learning applications.
The Physicists’ Toolbox for Machine Learning
Machine learning models can be viewed as simple: they take input data, process it within a computational "black box," and produce output. However, the complexity lies within that black box, specifically in the neural networks that approximate intricate functions.
Typically, programmers rely on trial and error to fine-tune these models, which can result in escalating costs. “As a trained physicist, I want to better understand what is happening inside these models to move away from trial and error,” Zhang expresses. “What are the properties that grant machine learning models their capabilities?”
Innovative Approaches to Scaling Laws
In his recent research published in the journal Machine Learning: Science and Technology, Zhang tackled the scaling laws of a particular model. His calculations required aggregating up to an infinite number of terms, a daunting task typical in physics.
To navigate this complexity, Zhang employed Feynman diagrams—a method developed by the physicist Richard Feynman in the 1940s for handling challenging quantum calculations. Instead of writing complicated equations, Feynman’s diagrams translate variables into visual representations, making it easier for the human brain to comprehend and track values involved in calculations.
Zhang applied this technique to a model discussed in earlier research, extending its analysis beyond previous limitations and deriving new scaling laws. His results promise to provide deeper insights into the model’s behavior and its potential applications.
The Human Element in AI Dynamics
As society dives into the world of AI, researchers, including Zhang, are focused on ensuring these tools are utilized responsibly. He raises an essential question about the consequences of AI technology: “We are building machines that may control us—it’s not about robots enslaving humans; it’s about how we struggle to understand the machines that profoundly influence our lives.”
The algorithms that govern social media, for example, can lead users down personalized, often isolating paths. Zhang’s concerns highlight the importance of understanding the ethical and societal implications of machine learning as we increasingly integrate AI into our daily lives.
Media & PR Contacts
For further information, you can reach out to:
- Lisa Potter
- Research Communications Specialist, University of Utah Communications
- Phone: 949-533-7899
- Email: lisa.potter@utah.edu