American researchers are working to introduce the laws of physics into machine learning models to improve the way algorithms understand the real world.
And if that wasn't unsettling enough, the effort is sponsored by DARPA – the spooky government agency primarily interested in defence matters.
The project is a collaboration between the University of Buffalo in New York and the Palo Alto Research Centre (PARC), an R&D subsidiary of Xerox.
"We're teaching physics to AI systems," chirped Rahul Rai, associate professor of mechanical and aerospace engineering in UB's School of Engineering and Applied Sciences.
"We are developing hybrid methods that integrate physics-based models – these are math-based formulas that explain the world around us, such as Einstein's E=MC2 – into the algorithms that guide machine learning, deep learning and other data-driven AI systems."
The team said this integration between physics and computer science will improve algorithms that deal with physically grounded systems – those originating in the real world – and help tackle "noisy data".
"Major limitations of existing purely data-driven statistical 'black box' methods include their inability to generalise beyond their initial set of training data, their agnostic view of underlying physics, resulting in model outputs that lack scientific coherency with the known laws of physics, and their 'data-hungry' nature that precludes them from being used in scientific problems and applications with limited or sparse data," PARC scientist Ion Matei explained.
Best known as a manufacturer of printers, Xerox has a long history of innovation: PARC, which it established in 1970, was responsible for such advances as the concept of the PC desktop, WYSIWYG text editors, Ethernet and the computer mouse. However, it was rarely successful commercialising the output of its labs.
In recent years, the company fell on hard times, and in 2018 Xerox was acquired by Japan's Fujifilm in a complex deal worth $6.1bn. ®