Earls, a mathematician and professor of civil and environmental engineering at Cornell University in Ithaca, N.Y., believes the key to improving the transparency of AI and putting humans fully in control of it is in applying more complicated math.
Beginning Sept. 1, Earls will lead Cornell University’s new Scientific Artificial Intelligence (SciAI) Center. The $11 million program, funded by a grant from the U.S. Office of Naval Research, focuses on establishing higher-level mathematics as the common language for AI.
Earls said the algorithms behind common AI applications today are largely based on algebra (focused on equations); involve less sophisticated concepts like weights, biases and pattern recognition; and are combined with massive sets of data strung together in simple building blocks to create billions of nodes that power machine-learning capabilities.
“At the end of it, it’s a black box, and you don’t know how it does what it does,” said Earls, explaining how the uncanny abilities of today’s AI that ‘spoof’ humans are at odds with true scientific discovery. “We want to pioneer new methods.”
If machine-learning models are powered by calculus (focused on rates of change) instead, AI could be better positioned to learn and understand how our world works similar to how humans do — through a direct step-by-step reasoning process that embraces the concept of cause and effect.
Instead of getting AI to predict the future using data from a physical system, we will get AI to speak in the language of calculus and derive the underlying differential equations that govern a physical system.
Alex Townsend, associate professor, Cornell University
“If AI is to learn about the universe, it will have to learn mathematics itself,” Earls said. “We already approach reasoning the same way. The idea is to have a human-machine partnership, rather than replacing humans.”
Alex Townsend, associate professor of math at Cornell, said this project continues what scientists have already been doing for centuries — modeling the world based on mathematical knowledge.
“Instead of getting AI to predict the future using data from a physical system, we will get AI to speak in the language of calculus and derive the underlying differential equations that govern a physical system,” Townsend said in a news release last month. “We are trying to develop an AI-human collaboration that can become our science teacher, revealing patterns of the natural world.”
While Cornell is the leader of this project, other institutions will play a role. Earls said he, Townsend and two other mathematicians from his school will work with peers from Brown, the U.S. Naval Academy, the University of Cambridge in the United Kingdom, and California universities Caltech, the University of California, Santa Cruz and the University of California, Berkeley. Graduate and post-doctoral students from all of those schools will be involved, and a nine-week immersion experience in Ithaca for all participants is being planned for next summer.
In the months and years to come, researchers with this project will identify shortcomings with existing data sets used in today’s AI, and work to develop scientific data sets capable of explaining cause and effect, Earls said. Obtaining more detailed and predictive explanations of how materials perform under stress is one example of how this new method of AI could be applied.
According to the news release, researchers at the SciAI Center will work on four functions — scientific data, user learning, complex systems and closure models.
“My two dreams are to completely revolutionize how we do science, and to use the tools from this center to understand complex, emerging behaviors,” Earls said.
