A team of researchers from the Intelligence Advanced Research Projects Activity and Lockheed Martin has started updating its dataset of satellite imagery to train their deep learning algorithms to identify indiscreet establishments, C4ISRNET reported Wednesday.
During the seven-month Functional Map of the World TopCoder challenge, IARPA and Lockheed developed artificial intelligence algorithms that were able to scan nuclear power plants, tunnel openings and wind farms, but failed to identify shipyards and ports, hospitals, office buildings and police stations.
Mark Pritt, a research scientist at Lockheed, explained that the algorithms might not recognize objects that lack distinguishing features.
He added that AI technologies are still not capable of copying the thinking and reasoning capabilities of humans.
Hakjae Kim, program manager of the fMoW challenge at IARPA, noted that IARPA and Lockheed are now working to prepare their algorithms for government use.
Researchers from Boston University are also leveraging the satellite imagery datasets and algorithms from the fMoW challenge to develop heat maps that show how algorithms classify objects.