Dr. Jason Matheny, director of the Intelligence Advanced Research Projects Activity (IARPA), gave the first keynote at USGIF’s Machine Learning and Artificial Intelligence Workshop Jan. 10 in Herndon, Va., with a strong articulation of why nearly 300 people were gathered for a day of presentations and discussion.
“We’re well past the point that it would’ve been possible for an agency of human analysts or 17 agencies of human analysts to have enough brains and eyeballs to perform their mission,” Matheny said. “So finding some way of bringing machines to scale up and address our problems is a way of bridging the gap between the resources we have and the resources we need. It allows us to focus human brains and eyeballs on the problems where they’re most needed.”
Machine learning (ML) progress accelerated in the last several years and has been especially celebrated in the case of deep learning, according to Matheny. He added vision is currently the most mature area of deep learning.
“Top performing object recognition systems have now risen to within a few percentage points of human accuracy and at substantially greater speed and lower cost than human labor,” he said. “Similarly, facial recognition systems have reached parity with humans.”
IARPA funds about 500 organizations around the world studying everything “from AI to Zika,” but is perhaps best known for its work in high-performance computing, Matheny continued. He pointed to IARPA’s many programs focused on automated analysis of speech, text, or video—such as Finder, CORE3D, Janus, Aladdin, and DIVA.
Matheny also said it’s important for the United States to monitor advancements in and uses of artificial intelligence (AI) and ML in other countries to avoid “technological surprise.”
IARPA often plays a role in AI and ML policy discussions, and co-led the recent White House research and development strategy for AI. As part of this effort, it was discovered that U.S. government investment in ML is about $1.2 billion per year, Matheny said.
“In contrast, industry’s research and development is at least 10 times larger and it’s growing at a much faster rate,” he elaborated. “ … So we need the private sector’s help in solving our hardest problems.”
Mo Islam, a senior associate with DFJ Venture Capital, presented following Matheny.
“The most valuable companies in the world currently are IT companies—and, I would argue, AI companies,” Islam said, referencing AI advancements by IT giants such as Google and IBM. “Every business is going to become an information business.”
Islam said he expects “edge computing,” the ability to conduct high-performance computing on handheld devices in low bandwidth areas, to be the next big thing in AI.
AI and the Third Offset
Dr. Nina Berry, a science and technology advisor with the Joint Improvised-Threat Defeat Organization (JIDO), gave a presentation discussing the relevance of AI and ML for the warfighter. She highlighted JIDO’s video intelligence (VIDINT) framework, which helps analysts process video to identify individuals.
“The AI we’ve been doing so far in VIDINT is showing great promise,” Berry said, adding the initiative supports the Pentagon’s Third Offset Strategy and its goal to implement more automation technology.
Dr. Catherine Cotell, the National Geospatial-Intelligence Agency’s (NGA) research anticipatory analytics lead, also discussed AI and the Third Offset Strategy. Cotell said in the face of a “tsunami” of commercial and small sat data NGA is “highly motivated” to automate not just the collection of the images but also the interpretation of them.
“If we don’t start doing some of this image processing with some automation we’re never going to be able to handle those massive amounts of data,” Cotell said.
She added that information storage is critical in addition to processing.
“None of those observations, none of that indication of change is going to be useful to anybody if it isn’t stored in an accessible way,” Cotell continued.
Dr. Peter Highnam, director of NGA Research, also spoke.
“Of all the INTs, GEOINT is where some of the AI tools really have a chance to shine,” Highnam said.
However, Highnam continued, AI and ML have a long academic history and ML’s performance is often misunderstood. In an effort to address this challenge, he said, the Defense Advanced Research Projects Agency (DARPA) has created a program called Explainable Artificial Intelligence (XAI), which aims to create a suite of ML techniques that produce more explainable models and enable users to understand, trust, and manage AI.
The workshop also featured academic and industry flash talks from a variety of universities and companies, an In-Q-Tel brief on natural language processing, and an update on the collaborative SpaceNet Challenge.
Following the workshop, attendees were invited to stay for USGIF’s bimonthly GEOINTeraction Tuesday networking event featuring Dr. David Honey of the Office of the Director of National Intelligence.