HRL Laboratories, LLC, has launched an ambitious project to develop a machine learning system that leverages knowledge of a set of labeled data onto a new unlabeled data set the way a child can recognize a zebra while only having ever seen a horse, but being told that zebras look like striped horses.
This area of research, zero-shot learning, motivates the program Automated Multisource Adaptation via Zero-shot Information Generation or AMAZING. The AMAZING project is being funded under the Defense Advanced Research Project Agency (DARPA)’s Learning with Less Labels (LwLL) program.
“The challenge we have with current machine learning methods is that they need millions of labeled data,” said HRL program manager Heiko Hoffmann. “This means that all these huge amounts of data gathered must be labeled for a computer to learn what the data are—an extremely costly process, especially for government agencies. The goal of the AMAZING system is to eliminate that effort with a system that enables a computerized intelligence to recognize data in a new set based on its knowledge of a previous set.” Soheil Kolouri is HRL’s principal investigator on the project.
In an extremely high-dimensional image space, with potentially millions of dimensions, it is challenging to find distributions of data and matches with previous data. The AMAZING technology enables automatic decomposition of image data, which makes it possible to match distributions between new and previous data sets. So new objects can be recognized based on their known components.
“I might have images or audio samples for which I have no labels at all,” Hoffmann said. “Our AMAZING machine or autonomous system will be able to learn what an object looks like based on previously learned models. Then, the AMAZING machine can leverage that knowledge onto a new data set with 90% recognition without labels, a rate that can be increased further with a few machine-selected labels. An autonomous system will be able to recognize new objects it’s never seen before, adding a new human-like layer of autonomy to the machine intelligence.”
Joining HRL on the AMAZING project are subcontractors Johns Hopkins University and University of Maryland, Baltimore County.
HRL Laboratories, LLC, Malibu, California (hrl.com) is a corporate research-and-development laboratory owned by The Boeing Company and General Motors specializing in research into sensors and materials, information and systems sciences, applied electromagnetics, and microelectronics. HRL provides custom research and development and performs additional R&D contract services for its LLC member companies, the U.S. government, and other commercial companies.
Media Inquiries: media[at]hrl.com