US War Department urgently needs to rethink machine learning strategy

“Tell the sovereign that they don’t clean their guns with the British: let them not clean them with us, otherwise, God bless the war, they are not good to shoot,” Levsha distinctly said, crossed himself and died.

A public report , prepared at Harvard in July 2017, shows how unprepared the U.S. military is in the field of artificial intelligence, such as deep learning.


Recently it became known that Google (the corporation of goodness) is engaged in the analysis of video images from military drones. This project is called Project Maven and was proposed in April 2017. Interestingly, collaboration with Google on this project was organized by Eric Schmidt himself, the former chairman of the Alphabet board of directors, and the current chairman of the DIB defense innovation board.

As part of the project, by the end of 2017 it was planned to introduce advanced algorithms on state computer platforms designed to highlight objects of interest in 38 classes in photo and video images. Colonel Zukor, who introduced the project, described the algorithm as 75 lines of code in Python.

Greg Allen, the author of the aforementioned Harvard report, tell your insider opinion that before the Maven project, no one in the Department had any idea how to properly acquire, implement, and test artificial intelligence. Let me emphasize this. A year ago, no one in the US Department of Defense had any idea about the above.

In fairness, we note that in DARPA there are several people who have an idea of ​​deep learning. In 2017, DARPA financed a number of modern projects: “Context Adaptation” , “Explainability” , “Biologically Inspired Architecture” .

There is clear evidence that deep learning is not taken seriously in the US defense industry. Here is a job descriptionpublished March 15, 2018. The U.S. Army is seeking a PhD in deep education with a gorgeous salary of $ 52,000. Maybe they forgot a zero?

F-35 aircraft costs at least $ 300 million. In comparison, Google acquired DeepMind for 500 million. But from the point of view of modern progress in deep training, the F-35 looks like the same antique military system as the cavalry on the eve of the Second World War. F-35 is so expensive, apparently because it is very complex. However, a swarm of 10,000 drones at a price of 1/300 of the cost of one F-35 will probably be just as effective on the battlefield.

This is precisely the problem of preparing for the wars of the future based on the technologies of the previous era. We cannot continue to spend 300 million each on an infinitely complex system that runs the risk of becoming obsolete on the very first day of arming against systems inspired by artificial intelligence.