Fighting Drone Swarms With Game Theory: The Many Ways AI Funding Can Advance U.S. National Security

Last month, the National Security Commission on AI issued its first recommendations to Congress, covering seven lines of effort, six of which are public and one of which is classified.  

I spoke with the commissioners leading the unclassified groups about their recommendations, beginning with Andrew Moore whose team considered the need to Increase AI R&D funding.

Andrew has a PhD from Cambridge University and did a postdoc at MIT before joining the faculty of Carnegie Mellon University. He is now Vice President of Google Cloud AI.

CRAIG:

The first quarter recommendations begin with a very specific and very ominous mention of swarm technology. Then your section recommends funding to understand how AI systems interact, including the application of game theory. As a layman reading that, I imagine dogfights between drone swarms. What are you envisioning in those references?

ANDREW:

Game theory refers to this whole area pioneered by John Nash, who was the subject of the movie, A Beautiful Mind. It addresses the question of what a person or a robot should do when they’re dealing with other people or robots who might have different goals. 

A perfect example that we all use in our regular lives is when you’re buying a house or a car or something like that. You don’t walk in and immediately tell the sales person exactly how much you’re willing to pay. You’re careful about how much information you divulge. Making artificial intelligences aware of the need to sometimes show discretion, not tip their hat in their activities, actually matters, certainly in things like national security or military operations.

CRAIG:

The recommendations talk about funding. Can you walk us through those recommendations in R&D funding for AI? There seem to be several lines of funding: defense funding, non-defense funding, and it would be helpful to understand that.

ANDREW:

The commission is recommending that we roughly double the funding for non-defense research spending from about a billion dollars a year to $2 billion a year in the United States. 

This will enable us to have the first mover technology in all the sub areas of artificial intelligence, ranging from things like understanding what is seen on a video screen through to planning the movements of tens of thousands or sometimes millions of different objects in a large operation, through to areas like anticancer technologies or general pandemic management systems. We want this investment used to increase the number of people in the United States, especially US citizens, who are trained to be in the AI workforce. 

We have armies of scientists trying out all kinds of experiments and, if you combine them with the power of artificial intelligence modeling, prediction and intelligence-guided experimentation, you get a multiplying effect.

The National Science Foundation will also get funding to work on the hardware, things like semiconductors, microelectronics and micro sensors that will be embedded all around the world to help get a better situational understanding of what’s going on. 

Meanwhile, NIST, the National Institute of Standards and Technology, is going to be extremely important as we wrestle with societal questions about how we would like to see artificial intelligences behave.

NASA is another important agency, which I would count as the number one institution on the planet for autonomy.

CRAIG:

You talk about enabling NIST to set standards, for example, in autonomous vehicles. What kind of testing standards do you imagine? Because it’s a very broad and fuzzy area, but the actual testing and the standards have to be fairly specific.

ANDREW:

That’s exactly right. So unfortunately, testing whether this system is safer than another system is very hard. You need sophisticated expertise of which the United States, in my opinion, has always been in the lead. The National Highway Safety Administration for the design of roadways has the technological expertise for building up testing frameworks.

CRAIG:

In the recommendations about NASA and the need for autonomous operations, you use the word ‘urgent.’ You call it an ‘urgent need.’ Why do you consider that urgent? Is there a particular threat that listeners should be aware of?

ANDREW:

The commission has been considering some specific scenarios which will go into a classified version of the report where there’s a particular need for urgency. So, the answer is yes, there are specific reasons, but outside of those specific classified reasons, whoever’s fastest in AI in 2020 is going to have an extra advantage in 2021 which leads to an extra, extra advantage in 2022

CRAIG:

You mentioned the strengthening fellowships and scholarships for domestic students. Is there a concern that the US is becoming overly reliant on international students? 

ANDREW:

It’s a very complex question. First, we agree that it’s absolutely essential that we grow the proportion of US citizens who are trained in these technologies who can work on federal defense and national security problems. 

Second, we all agree that the United States has attained a huge technological advantage over the last half century is because the best minds from all around the world come and join in.

We need to make sure that we remain the world’s center of artificial intelligence and we need to make sure that there is still a large proportion who are domestic United States citizens. 

CRAIG:

The recommendations include funding for unsupervised, self-supervised and transfer learning. That’s basic research and the recommendations in another section talks about a five to one ratio of government funding for applied versus basic research. What do you regard as a healthier ratio and are the recommendations intended to right that balance somewhat?

ANDREW:

The five to one ratio in the report was specifically referring to hardware and hardware systems. It varies across different parts of technology. 

Let’s suppose, for instance, that we want by 2028 to have autonomous medical recovery helicopters based on electric helicopter motor technology. That’s a very practical problem. It doesn’t sound like basic research.

But, I can guarantee you that’s once it’s out there in the field, that helicopter will see a situation that no helicopter has seen before. Transfer learning is all about autonomous systems which have learned how to do things one way, recognizing a new situation and having to relearn on the spot based on lessons learned in earlier stages of training. We need to see some very significant progress in basic research on transfer learning before we will be able to build on autonomous helicopter for a medical evacuation.

The way that modern research is done is that you have people who are doing the basic research and practicing on really hairy applied problems like the one I just described. So, we see less compartmentalization between the different stages of research. 

CRAIG:

Further on in the recommendations, you talk about a national AI research resource that would democratize AI by making resource intensive computation and large curated data sets available to researchers. Can you talk about how that would work? 

ANDREW:

Commercial researchers inside companies usually have access to all kinds of important useful data to test their big ideas on. Someone inside a maps-based company can experiment with two or three orders of magnitude more data than the datasets available for academic researchers to use. That can lead to brilliant students who would have done something great having to play with toy problems. The last thing we want is our best young minds just playing with toy problems. We need the academic United States to share and experiment with data at a much larger level.

The researcher should worry about research and not questions of trying to chase down links to lots of distributed data sets everywhere. I’ve been involved in plenty of research projects where you have two years to solve a big problem and you suddenly realize you’re 15 months into it and you’re still hunting down the data.

If we can get that right so that the researcher can spend all their time doing the actual science, we will have made sure that some of our most brilliant minds in the country are being much more productive and able to be much more creative about what they’re doing with that data.

CRAIG:

I had a conversation with Eric Schmidt and Bob Work a few months ago and they talked about data sharing among allies. Presumably this resource would not only be connected to domestic data sources, but among allies as well. 

ANDREW:

In every form of data, you’re going to see different access control levels. You’d never have a binary switch where this data is completely available to the entire planet vs this data is only available to this one authorized user with a password. And so, building these large data stores and making them convenient to use has many facets. One of them is setting up the access control policies so that the right people have access to data at the right level. 

As we talk about sharing with allies, I see that as part of this bigger picture of any data management system and an inevitable part of the design of that system is to make sure that people with the right level of authority can only access the data that they should.

CRAIG:

Do you have a sense if these recommendations are implemented and presumably recommendations to follow, is the US going to leap ahead?

ANDREW:

For me as a commissioner, one of the things I most want to see is the United States get to a point where we have a technologically nimble workforce of millions of people throughout the services, throughout the government and the private sector who are able to very rapidly use technology to solve immediate problems in automation. And that’s what we’re going to get out of this increased level of funding.