Scientists can learn a physical system from one experiment, math shows

Some experiments provide only one chance to gather data, but that one shot can be enough for scientists to learn the equations governing a physical system, Cornell researchers have found. They offer a mathematical foundation for learning physical models from fewer experiments. 

The team has proven with an all-or-nothing result – known as a zero-one law – that for a broad class of machine learning models, if any single experiment can identify the model, then almost any randomized experiment will succeed. 

“This result changes how we think about designing experiments,” said co-author Alex Townsend, associate professor of mathematics and Stephen H. Weiss Junior Fellow in the College of Arts and Sciences (A&S). “Usually, we imagine that we need to carefully choose the right experiment to learn how a system works. But our theory says that, for a broad class of models, if one experiment can reveal the underlying equations, then almost any randomly chosen experiment does the trick. 

“To me, that’s a remarkable simplification in our thinking. We don’t need to find the perfect experiment; instead, we can pick any old one at random.” 

“A zero-one law for one-shot system identification” published in Proceedings of the National Academy of Sciences on Nov. 20. 

Townsend and Samuel Otto, assistant professor in the Sibley School of Mechanical and Aerospace Engineering, in the Cornell Duffield College of Engineering, are co-authors with collaborators from Imperial College, London, and Courant Institute at New York University. 

The findings are a useful insight for experiments that are costly, destructive or hard to repeat, Otto said. One-shot learning could also be important when a system abruptly changes and a model for its new behavior needs to be determined quickly, Otto said, such as a mechanical system that must still be safely operated after the failure of a component. 

Creating a machine learning model based on data helps experimenters predict what the system will do with different data inputs, Otto said. Just how much – or how little – data is needed to create a reliable model is a fundamental question. 

The researchers examined a hypothetical system, set up as a computer simulation that might, for instance, measure what happens when a force is applied over time. 

“We started observing that in our numerical experiments, depending on how the system was set up, there was an all or nothing result,” Otto said. 

Either the system parameters could be identified with only one data input, or it would be impossible to identify the system parameters applied because of a problem fundamental to the given setup. 

“There was nothing in between,” Otto said. “It was 100% or 0%.” 

They proved theoretically that the set of unsuccessful inputs, in which the parameters cannot be determined from the resulting output, is either infinitesimally small or it’s the only possibility. There’s no middle ground, such as inputs that work 30% of the time. 

Otto pictures a dart board. “If you throw a dart, your set of unsuccessful inputs is either the infinitesimally thin lines that separate the scores on the dart board – you’ll never hit it – or it’s the whole board,” he said. 

This type of machine learning model is used in a wide range of research, so the result is broadly applicable, Otto said. It can be used to learn systems that evolve over time, as well as mechanical systems – for instance, how various parts of a bridge are displaced under the force of a train crossing. In chemistry, the result could be used in a reaction diffusion equation, the mathematical model that describes how two or more chemicals react to each other and diffuse through a medium over time. 

“For this type of model, we were able to show that a small amount of data is sufficient,” Otto said. 

The researchers demonstrate the concept in the paper with simulations of fluid flow and chaotic motion. 

There are limitations, however. The result applies to a specific type of machine learning model, which is linear with respect to its parameters, Otto said, as opposed to a neural network. 

It’s also correct only in a theoretical, “noiseless” setting, in which both the input and the output can be measured entirely and perfectly. 

“Obviously, the world is noisy. The world is not completely observable,” Otto said. “You can’t, for instance, observe the displacement of a bridge at every point on the bridge.” 

They have work in progress about extending the result in a noisy setting, he said. 

“In the noiseless case, you can think of the set of unsuccessful inputs as infinitesimally thin,” he said, referring to the imagined dart board. “Noise makes that set a little bit thicker. We’re trying to determine how much thicker the lines get.” 

The study was supported by the National Science Foundation and the U.S. Department of Defense Advanced Research Projects Agency (DARPA). 

Kate Blackwood is a writer for the College of Arts and Sciences.

Media Contact

Adam Allington