Rose Yu’s plan to make AI faster, smarter and better is already showing results.
Rose Yu builds neural networks using the principles of physics at the University of California San Diego. She is looking over La Jolla Shores.
Peggy Peattie,How much magazine.
Introduction
She was 10 years old. Rose yu received a birthday gift that could change her life and the way we learn physics. Her uncle gave her a computer. It was a very rare gift in China 25 years earlier, and it wasn’t wasted. Yu played computer games at first, but she won a web design award in middle school. It was the first in a series of computer-related awards.
Yu majored in computer science at Zhejiang University where she won a research prize for innovative work. She chose the University of Southern California for her graduate studies partly because she knew only one person in the United States, and that was her uncle who worked at the Jet Propulsion Laboratory near Pasadena. Yu received her doctorate in 2017, and was awarded for the best dissertation. In January, President Joe Biden presented her with the Presidential Early Career Award in his final week as president.
Yu is now an associate professor at University of California, San Diego. She has spent years integrating our knowledge of physics with artificial neural networks. Her work has not only led to novel techniques for training and building these systems, but also helped her make progress with several real-world applications. She has used fluid dynamics principles to improve traffic forecasts, accelerated simulations of turbulence in order to better understand hurricanes, and devised tools to help predict the spread Covid-19.
Yu’s work has brought her closer to her grand vision — deploying a set of digital lab assistants she calls AI Scientist. She envisions a “partnership”based on physics principles, between human researchers and AI-based tools that can yield new scientific insights. She believes that combining the inputs of a team with such assistants may be the best method to boost the discovery process.
Quanta ( ) spoke with Yu on turbulence, its many faces, and how AI can help us get out of gridlock in cities. The interview has been edited and condensed for clarity.
Yu on the UCSD campus, where she is an associate professor.
Peggy Peattie forHow much magazine
When was the first time you tried to combine physics and deep learning?
The problem began with traffic. I was a graduate student at USC. The campus is located near the intersection of I-10 & I-110. Traffic can be annoying when you are trying to get anywhere. In 2016, I started to wonder if I could do something about it.
Back then, deep learning — which uses multilayered networks to elicit data patterns — was a hot topic. There was a lot excitement around applications in image classifying, but images are static. I wondered if deep learning could be used to solve problems where the situation is constantly changing. I wasn’t first to think about this, but I and my colleagues came up with a new way to frame the problem. What was your new method?
We first thought of traffic as a physical process of diffusion. In our model, traffic flow over a road network is analogous to fluids flowing over a surface – motions that are governed the laws of fluid dynamic. Our main innovation was to view traffic as a mathematical graph. This is based on the field of graph theory. Sensors that monitor traffic on highways or other roads serve as nodes in this graph. The edges of the graph are the roads (and distances between them) that these sensors monitor.
Yu’s interest in computers began with a gift for her 10th birthday.
Peggy Peattie forHow much magazine
This graph gives you a snapshot of all the roads at a certain time. It also tells you the average speed of cars at each point on the graph. You can get a good idea of traffic flow by combining a series of snapshots taken every five minutes. You can then try to predict the future.
In deep learning, you need a large amount of data to train your neural network. Cyrus Shahabi had been working on traffic forecasting for many years and had a large amount of data from L.A. that I could access.
How accurate were your predictions?
Before our work, people were only able to make traffic forecasts which were reliable for 15 minutes. Our forecasts were accurate for an hour, a significant improvement. Google Maps used our code in 2018. Google invited me as a visiting researcher a few months later.
Is that when you started working on climate modelling?
I did give a talk in 2018 at the Lawrence Berkeley National Laboratory. After that, I met with scientists at the Lawrence Berkeley National Laboratory and we discussed a problem which would be a good testing ground for physics guided deep learning. We chose to predict the evolution of turbulent flows, which is an important factor in climate models and a major area of uncertainty.
Turbulence is best illustrated by the swirling patterns that appear after you pour milk into a coffee cup and stir it. These swirls can cover thousands of miles in the oceans. The Navier-Stokes Equation, which describes fluid flow, is used to predict turbulent behavior. This equation is considered the gold standard for this field. The calculations required are slow, and this is why we do not have good models to predict hurricanes and tropical storms.
The heavy congestion of Los Angeles first inspired Yu to model highway traffic as the flow of fluids.
Peggy Peattie forHow much magazine
Can deep learning help?
Deep neural networks trained on our best numerical simulators can learn how to “emulate”or imitate, those simulations. They do this by recognizing patterns and properties hidden within the data. They don’t need to do time-consuming brute-force calculations in order to find approximate solutions. Our models increased the speed of predictions by a factor 20 in two-dimensional settings, and by a multiplier 1,000 in three-dimensional ones. Someday, a module like our turbulence-prediction module could be inserted into larger climate models to improve their ability to predict things like hurricanes.
What other places does turbulence appear?
It is everywhere. Turbulence of blood flow can cause heart attacks or strokes. When I was a Caltech postdoc [California Institute of Technology]I co-authored an article that looked at stabilizing drones. Airflows generated by propellers interact with the earth to create turbulence. This can cause the drones to wobble. We used a neuronal network to model turbulence and this led to better control for drones during takeoffs, landings, and other operations.
Currently, I’m working with scientists from UCSD and General Atomics to develop fusion power. The key to success is to learn how to control plasma, a hot, ionized state of matter. Different types of turbulence occur in the plasma at temperatures of around 100 million degrees. The numerical models based on physics that describe this behavior are extremely slow. We are developing a deep-learning model that will be able predict the behavior of plasma in a split-second, but it is still in development.
Yu and doctoral student Jianke Yang in her office at UCSD.
Peggy Peattie forHow much magazine
Over the last two years, my team has developed AI algorithms which can automatically discover symmetry principle from data. Our algorithm, for example, identified the Lorentz-symmetry, which is related to the constancy of light speed. Our algorithm also identified rotationalsymmetry, which is the fact that a sphere does not look different no matter how you rotate it. This was something it wasn’t specifically trained to know. Our tools have the ability to discover new symmetries that are not known to physics. This would be a major breakthrough.
I then thought, if our tools are able to discover symmetries using raw data, why not generalize it? These tools could also be used to generate new research ideas or hypotheses for science. This was the origin of AI Scientist.
Is AI Scientist just a fancy version of a neural network?
This is not a single network but an ensemble of computer programmes that can help scientists discover new things. My group has developed algorithms to help with specific tasks, like weather forecasting, identifying drivers of global temperatures rise, or trying discover causal relationships, such as the effects of vaccination policy on disease transmission.
Now, we’re building a “foundation” that can handle multiple tasks. Scientists collect data from a variety instruments. We want our model to be able to handle a variety of data formats, including text, images, and videos. We have a prototype, but want to make it more comprehensive, intelligent, and better trained. This could happen in a few years. What do you think it could do
?
AI is able to assist in virtually every step of the discovery process. When I refer to an “AI Scientist”I mean a scientific assistant. For example, the literature survey stage of an experiment requires a massive amount of data collection and organization. A large language model is now able to read and summarize thousands books in a single lunch hour. AI is not very good at judging scientific validity. In this case, AI cannot compete with an experienced researcher. AI can help with hypothesis generation, design of experiments, and data analysis but it cannot carry out sophisticated experiment.
What is the maximum you would like to see this concept develop?
In my mind, an AI Scientist would relieve researchers of some of their drudgery and let people focus on the creative aspects of scientific research. We’re very good at that. The goal is not to replace scientists. I can’t imagine — nor would I want to ever see — a machine replacing, or interfering, with human creativity.
Next article
How the strange shape of a biofilm is derived from its cellular geometry
