Opening the Right Doors: Konstantinos Spiliopoulos on the Future of Scientific AI
Cruising at 30,000 feet, the plane suddenly jolts. The “Fasten Seatbelt” sign dings, and for a few seconds, the ride is bumpy.
To a passenger, this is an annoying phenomenon, but to a scientist, it’s one of the unsolved problems of physics— turbulence.
There are equations to describe the flow of air, but they can be so complex that even supercomputers can’t predict how the air will move just seconds into the future.
Unpredictabilities of that sort are exactly what CISE Faculty Affiliate Konstantinos Spiliopoulos, Professor and Director of Statistics at the Department of Mathematics and Statistics, is trying to solve.
“One of the unifying themes between statistics, probability, applied mathematics, and machine learning,” Spiliopoulos explained, “is trying to model and quantify the effect of randomness on different phenomena.”
Instead of trying to fix the chaos, we should be treating it as a ‘structured’ kind of randomness, finding clear patterns in what looks like total noise to the outside world.
Large Language Models (LLMs) are large, expensive, and consume enough energy to power small cities. Spiliopoulos compared how these models currently function to the 100 Doors Problem.
“Imagine you are entering a room that has 100 doors,” he said. “A couple of them take you to the right answer, but you don’t know which ones. Currently, models try to open and close all 100 of them until they find the right combination.”
Corporations have massive “compute” power, so they can afford to open every door. They find the patterns, but they do it inefficiently. Spiliopoulos’ goal is to mathematically determine which doors are essential and which are just “noise” that can be avoided.
“In certain cases, we can sacrifice a bit of accuracy for a tremendous gain in computational complexity,” he noted.
By accepting a tiny margin of error, Spiliopoulos’s work focuses on using statistical and probabilistic tools to reduce the amount of processing power a model needs. This makes AI much more efficient, allowing it to run faster while using a fraction of the electricity required by today’s ‘brute force’ methods.
Moving away from this power-hungry approach is becoming a necessity as the digital and physical costs of AI continue to climb. The recent surge in AI popularity has created a strange paradox on college campuses.
More students are using these tools than ever, but fewer understand how it actually works. Spiliopoulos noticed this gap back in 2014. “There was a gap in the training of students,” he recalled. “Many students focus on the computational aspect, but they don’t necessarily know why it is working.”
This realization led to a new course at BU and, eventually, a formal textbook, Mathematical Foundations of Deep Learning Models and Algorithms.
For Spiliopoulos, it wasn’t just about teaching code, but about teaching the “why.”
“Without understanding the mathematical foundations, we are simply users,” he noted. The textbook serves as a roadmap for the next generation of researchers to understand the statistical mechanics of the tools they use every day.
Spiliopoulos offered a sobering but exciting reality check for the technological future. He believes we are reaching a phase of the curve where “brute force” pattern recognition won’t be sustainable at the current pace of progress.
To solve the truly hard questions, AI needs to do more than just guess the next word in a sentence. It needs to respect the laws of nature.
“Nature works in a different way,” he explained. “It’s not just about patterns. We don’t fully understand the physical or biological laws yet.”
He added that the field is shifting toward “Scientific AI,” where models are built to understand the fundamental laws of physics and biology. This shift requires a new kind of student: one who isn’t just a “user” of AI, but an innovator.
“Education will have to move more towards adopting these tools in a way that makes us smarter, not necessarily just more productive,” Spiliopoulos concluded.
The goal in the search to find the singular correct door in a row of 100 is not just to find the answer, but to understand why that door was the right one in the first place. For Spiliopoulos, it isn’t a search for efficiency, but clarity.
By stripping away the “noise” and focusing on the underlying mathematics, he is helping usher in an era in which AI is a precision tool rooted in the laws of nature.
Konstantinos Spiliopoulos is a Professor and Director of Statistics at the Department of Mathematics and Statistics at Boston University. He is also a Faculty Affiliate of the Center for Information and Systems Engineering, and has been a junior faculty fellow at the Hariri Institute for Computing and Computational Sciences & Engineering at Boston University.
His research interests include stochastic processes; algorithmic and computational methods in machine and statistical learning; applied mathematics and probability; financial mathematics; asymptotic problems for stochastic processes; partial differential equations, such as multiscale methods and large deviations; and statistical analysis and inference.