21 December 2017

Project Maven brings AI to the fight against ISIS

Gregory C. Allen

Gregory C. Allen

Gregory C. Allen is an adjunct fellow at the Center for a New American Security. He is the co-author of “...


Maven is designed to be that pilot project, that pathfinder, that spark that kindles the flame front of artificial intelligence across the rest of the [Defense] Department.

Air Force Lt. Gen. Jack Shanahan, November 2017

For years, the Defense Department’s most senior leadership has lamented the fact that US military and spy agencies, where artificial intelligence (AI) technology is concerned, lag far behind state-of-the-art commercial technology. Though US companies and universities lead the world in advanced AI research and commercialization, the US military still performs many activities in a style that would be familiar to the military of World War II.

As of this month, however, that has begun to change. Project Maven is a crash Defense Department program that was designed to deliver AI technologies—specifically, technologies that involve deep learning neural networks—to an active combat theater within six months from when the project received funding. Most defense acquisition programs take years or even decades to reach the battlefield, but technologies developed through Project Maven have already been successfully deployed in the fight against ISIS. Despite their rapid development and deployment, these technologies are getting strong praise from their military intelligence users. For the US national security community, Project Maven’s frankly incredible success foreshadows enormous opportunities ahead—as well as enormous organizational, ethical, and strategic challenges.

Founding Project Maven. In late April, Robert Work—then the deputy secretary of the Defense Department—wrote a memo establishing the Algorithmic Warfare Cross-Functional Team, also known as Project Maven. The team had only six members to start with, but its small size belied the significance of its charter. Project Maven—directed by Air Force Lt. Gen. Jack Shanahan and led by Marine Corps Col. Drew Cukor—was tasked with developing and fielding the first operational use of deep learning AI technologies in the defense intelligence enterprise. The Defense Department has long funded basic research and development in AI and has fielded semi-autonomous systems. But Project Maven is the first time the Defense Department has sought to deploy deep learning and neural networks, at the level of state-of-the-art commercial AI, in department operations in a combat theater.

Before Project Maven’s creation, the Defense Department was advised by leading AI experts in industry and academia to seek out a narrowly defined, data-intensive problem where human lives weren’t at stake and occasional failures wouldn’t be disastrous. Fortunately for the team, the defense intelligence community is currently drowning in data. Every day, US spy planes and satellites collect more raw data than the Defense Department could analyze even if its whole workforce spent their entire lives on it.

As its AI beachhead, the department chose Project Maven, which focuses on analysis of full-motion video data from tactical aerial drone platforms such as the ScanEagle and medium-altitude platforms such as the MQ-1C Gray Eagle and the MQ-9 Reaper. These drone platforms and their full-motion video sensors play a major role in the conflict against ISIS across the globe. The tactical and medium-altitude video sensors of the Scan Eagle, MQ-1C, and MQ-9 produce imagery that more or less resembles what you see on Google Earth. A single drone with these sensors produces many terabytes of data every day. Before AI was incorporated into analysis of this data, it took a team of analysts working 24 hours a day to exploit only a fraction of one drone’s sensor data.

The Defense Department spent tens of billions of dollars developing and fielding these sensors and platforms, and the capabilities they offer are remarkable. Whenever a roadside bomb detonates in Iraq, the analysts can simply rewind the video feed to watch who planted it there, when they planted it, where they came from, and where they went. Unfortunately, most of the imagery analysis involves tedious work—people look at screens to count cars, individuals, or activities, and then type their counts into a PowerPoint presentation or Excel spreadsheet. Worse, most of the sensor data just disappears—it’s never looked at—even though the department has been hiring analysts as fast as it can for years.

Thousands of people in the department currently work on analyzing full-motion drone video data. Plenty of higher-value analysis work will be available for these service members and contractors once low-level counting activity is fully automated. As such, Project Maven won’t exactly pay for itself through savings on salaries. Nevertheless, the benefit of automating this specific task—as well as the benefit that other Defense Department projects will derive from leveraging Maven’s AI capabilities and infrastructure—mean that Project Maven has more than justified its price tag of about $70 million.

The Project Maven playbook. The six founding members of Project Maven, though they were assigned to run an AI project, were not experts in AI or even computer science. Rather, their first task was building partnerships, both with AI experts in industry and academia and with the Defense Department’s communities of drone sensor analysts.

In the political aftermath of the Edward Snowden leaks and Donald Trump’s election, tech companies have been wary of helping the national security community address its tech challenges. AI experts and organizations who are interested in helping the US national security mission often find that the department’s contracting procedures are so slow, costly, and painful that they just don’t want to bother. Project Maven’s team—with the help of Defense Information Unit Experimental, an organization set up to accelerate the department’s adoption of commercial technologies—managed to attract the support of some of the top talent in the AI field (the vast majority of which lies outside the traditional defense contracting base). Figuring out how to effectively engage the tech sector on a project basis is itself a remarkable achievement.

Access to the right talent and partnerships allowed Project Maven to structure its program correctly from the outset. Before Maven, nobody in the department had a clue how to properly buy, field, and implement AI. A traditional defense acquisition process lasts multiple years, with separate organizations defining the functions that acquisitions must perform, or handling technology development, production, or operational deployment. Each of these organizations must complete its activities before results are handed off to the next organization. When it comes to digital technologies, this approach often results in systems that perform poorly and are obsolete even before they are fielded.

Project Maven has taken a different approach, one modeled after project management techniques in the commercial tech sector: Product prototypes and underlying infrastructure are developed iteratively, and tested by the user community on an ongoing basis. Developers can tailor their solutions to end-user needs, and end users can prepare their organizations to make rapid and effective use of AI capabilities. Key activities in AI system development—labeling data, developing AI-computational infrastructure, developing and integrating neural net algorithms, and receiving user feedback—are all run iteratively and in parallel.

Though modern AI techniques for imagery analysis are extremely capable, developing algorithms for a specific application is not yet effortless—not just plug-and-play. Building robust, deep learning AI systems requires huge data sets with which to train the deep learning algorithm. Training data must not only be available, but categorized and labeled in advance by humans. Paradoxically, this phase of automation can be very labor-intensive. In Maven’s case, humans had to individually label more than 150,000 images in order to establish the first training data sets; the group hopes to have 1 million images in the training data set by the end of January. Such large training data sets are needed for ensuring robust performance across the huge diversity of possible operating conditions, including different altitudes, density of tracked objects, image resolution, view angles, and so on. Throughout the Defense Department, every AI successor to Project Maven will need a strategy for acquiring and labeling a large training data set.

Once labeled data is ready, the algorithmic training process makes extremely intensive computational demands. Traditional IT infrastructure is practically useless for such computations. Many leading commercial tech companies have gone so far as to develop their own custom processors and cloud infrastructure networks to run AI computations. The department has spent years, and billions of dollars, trying to migrate its digital activity into the cloud, but none of that infrastructure was built with requirements for AI training and inference computation in mind. Project Maven had to build its own AI-ready infrastructure, including computing clusters for graphics processing, from scratch. Fortunately, some of this capability can be leveraged for future algorithm training on other department projects.

Even before the final versions of Project Maven’s labeled data set and computational infrastructure were ready, the alpha and beta versions were used to develop algorithms that were shared with the user community to get feedback. Maven’s team heard from users with full-motion video know-how in the specific context of counter-ISIS operations in the Middle East. From their users, Maven’s developers found out quickly when they were headed down the wrong track—and could correct course. Only this approach could have provided a high-quality, field-ready capability in the six months between the start of the project’s funding and the operational use of its output. In early December, just over six months from the start of the project, Maven’s first algorithms were fielded to defense intelligence analysts to support real drone missions in the fight against ISIS.

A hundred Mavens? The good news is that Project Maven has delivered a game-changing AI capability. In doing so, the effort has demonstrated a level of technological innovation and programmatic agility that has been sorely lacking from most Defense Department digital initiatives. The bad news is that Project Maven’s success is clear proof that existing AI technology is ready to revolutionize many national security missions—even if the department is not yet ready for the organizational, ethical, and strategic implications of that revolution.

Now that Project Maven has met the sky-high expectations of the department’s former second-ranking official, its success will likely spawn a hundred copycats throughout the military and intelligence community. The department must ensure that these copycats actually replicate Project Maven’s secret sauce—which is not merely its focus on AI technology. The project’s success was enabled by its organizational structure: a small, operationally focused, cross-functional team that was empowered to develop external partnerships, leverage existing infrastructure and platforms, and engage with user communities iteratively during development. AI needs to be woven throughout the fabric of the Defense Department, and many existing department institutions will have to adopt project management structures similar to Maven’s if they are to run effective AI acquisition programs. Moreover, the department must develop concepts of operations to effectively use AI capabilities—and train its military officers and warfighters in effective use of these capabilities. As one astute strategist at the Pentagon told me, “The tech itself is of limited utility if we don’t also have people who can use it, ideas about how to use it, and training to be good at using it.”                                                               

Already the satellite imagery analysis community is working on its own version of Project Maven. Next up will be migrating drone imagery analysis beyond the campaign to defeat ISIS and into other segments of the Defense Department that use drone imagery platforms. After that, Project Maven copycats will likely be established for other types of sensor platforms and intelligence data, including analysis of radar, signals intelligence, and even digital document analysis. It won’t stop there, either. Lt. Gen. Shanahan is especially optimistic about AI capabilities in the cyber domain, and the National Security Agency agrees. In October 2016, Michael Rogers (head of both the agency and US Cyber Command) said “Artificial Intelligence and machine learning—I would argue—[are] foundational to the future of cybersecurity. … It is not the if, it’s only the when to me.”

The US national security community is right to pursue greater utilization of AI capabilities. The global security landscape—in which both Russia and China are racing to adapt AI for espionage and warfare—essentially demands this. Both Robert Work and former Google CEO Eric Schmidt have said that leadership in AI technology is critical to the future of economic and military power and that continued US leadership is far from guaranteed. Still, the Defense Department must explore this new technological landscape with a clear understanding of the risks involved.

To its credit, the department selected drone video footage as an AI beachhead because it wanted to avoid some of the more thorny ethical and strategic challenges associated with automation in warfare. As US military and intelligence agencies implement modern AI technology across a much more diverse set of missions, they will face wrenching strategic, ethical, and legal challenges—which Project Maven’s narrow focus helped it avoid. The stakes are relatively low when AI is merely counting the number of cars filmed by a drone camera, but drone surveillance data can also be used to determine whether an individual is directly engaging in hostilities and is thereby potentially subject to direct attack. As AI systems become more capable and are deployed across more applications, they will engender ever more difficult ethical and legal dilemmas.

US military and intelligence agencies will have to develop effective technological and organizational safeguards to ensure that Washington’s military use of AI is consistent with national values. They will have to do so in a way that retains the trust of elected officials, the American people, and Washington’s allies. The arms-race aspect of artificial intelligence certainly doesn’t make this task any easier. A recent news headline—“Russia to the United Nations: Don’t try to stop us from building killer robots”—is, sadly, not as hyperbolic as it may sound.

Beyond legal and ethical concerns, the vulnerabilities and failure modes of AI technology are quite different from those of traditional software. It is possible, for example—through use of image modifications undetectable to the human eye—to generate “adversarial examples” that reliably fool AI computer vision systems. AI systems are also vulnerable to cyber attacks that “poison” the training data set. The Defense Department must develop and field AI systems that are reliably safe when the stakes are life and death—and when adversaries are constantly seeking to find or create vulnerabilities in these systems.

Moreover, the department must develop a national security strategy that focuses on establishing US advantages even though, in the current global security environment, the ability to implement advanced AI algorithms diffuses quickly. When the department and its contractors developed stealth and precision-guided weapons technology in the 1970s, they laid the foundation for a monopoly, nearly four decades long, on technologies that essentially guaranteed victory in any non-nuclear war. By contrast, today’s best AI tech comes from commercial and academic communities that make much of their research freely available online. In any event, these communities are far removed from the Defense Department’s traditional technology circles. For now at least, the best AI research is still emerging from the United States and allied countries, but China’s national AI strategy, released in July, poses a credible challenge to US technology leadership.

Project Maven has successfully brought best-of-breed AI technology, along with project management practices from the commercial technology sector, to the US military. That is a remarkable achievement. But where artificial intelligence and national security are concerned, it is just the beginning of the challenges the United States will face.