Sneaking a Peek at DARPA's AI Colloquium
The research agency seeks major advances in artificial intelligence.
Artificial intelligence (AI) has come a long way in recent years, but the technology still has hurdles to overcome if machines are to become true partners and collaborators with humans. To help push the systems to that next level, the Defense Advanced Research Projects Agency (DARPA) is hosting a two-day conference aimed at spurring the next wave of AI advances.
The Artificial Intelligence Colloquium being held March 6-7 in Alexandria, Virginia, is designed to bring together AI researchers from across the Defense Department, industry and academia with the broader military community seeking to use AI technologies. Attendees will learn more about DARPA’s current and emerging AI programs and how the myriad technologies in development could apply to their diverse missions.
“We’re going to have 18 briefings in total. The goal of the colloquium is to go a bit deeper into the technology than we typically do. For many events, we try to appeal to a broad audience. For this one, it’s more of a technical audience,” explains John Everett, deputy director of DARPA’s Information Integration Office. “We’re going to try to give a view of recent advances. The breadth of AI across the agency ranges from cyber to chemistry and design and natural language processing. There’s an enormous range, so that means there’s an enormous range of performers.”
DARPA officials describe deep learning as the second wave of AI technologies, which still has major limitations. Agency officials aim to foster a third wave of technologies that will allow the systems to be trusted partners for humans, whether in the science lab, in cyberspace or on the battlefield. The agency is investing up to $2 billion in new and existing programs as part of its AI Next Campaign and has about 80 AI-related programs across the organization, including areas such as scientific discovery, cyber defense, software engineering, aviation and spectrum management.
The multitude of programs include:
AI Research Associate, which will attempt to “impart intuition” and an ability for the AI to actually recommend new scientific experiments or new avenues of research, explains Valerie Browning, who directs DARPA’s Defense Sciences Office.
Physics of AI, which aims to develop novel architectures, algorithms and approaches that include physics, mathematics and prior knowledge relevant to an application domain so that AI can be better applied in scientific discovery, human-AI collaboration, and a variety of defense applications. “As a physicist or as a scientist, there are certain fundamental thermodynamic laws that help us to exclude certain solutions. AI is not smart enough to do that yet, so we need to figure out how to embed that knowledge to help focus research toward practical solutions,” Browning says.
Explainable AI, which aims to create a suite of machine learning techniques that produce more explainable models while maintaining a high level of learning performance, and to enable human users to understand, appropriately trust and effectively manage the emerging generation of artificially intelligent partners.
Cyber-Hunting at Scale, also known as CHASE, which seeks to develop automated tools to detect and characterize novel attack vectors, collect the right contextual data, and disseminate protective measures both within and across enterprises.
And, Active Interpretation of Disparate Alternatives, a natural language tool for putting multimedia, such as text, speech, images, video and metadata, into a common representational format.
The colloquium is part of DARPA’s effort to dramatically push the state of the art rather than settle for incremental improvements. “Many people think of AI as machine learning, but there are many more techniques available,” Everett says. “We’re trying to show the range of the performer-base and the range of techniques we’re developing.”
For more information on DARPA’s artificial intelligence programs, be sure to read the March issue of SIGNAL Magazine.