Helping government leverage AI tools

Nick Weir, challenge manager for SpaceNet at In-Q-Tel's CosmiQ Works lab, discusses how his project crowdsources AI solutions for government applications.

Artificial intelligence has already found a wide range of useful applications, but there’s always more that can be done with it and technologies like it. One of the companies working hard to help the government, non-profits, and commercial businesses realize its potential is In-Q-Tel, through their lab CosmiQ Works. To learn more about how the lab is assisting these fields, we spoke with Nick Weir, In-Q-Tel Data Scientist and challenge manager for their SpaceNet project.

ABERMAN: Well Nick, I put you on the spot, but what is SpaceNet?

WEIR: SpaceNet is a nonprofit partnership, single member led LLC led by CosmiQ Works, which as you mentioned is a lab at In-Q-Tel. And we’re a partnership along with Maxar Technologies, Amazon Web Services, and the Intel AI Lab. And our main effort is to produce datasets of overhead imagery, which includes both the images and then labels for mapping purposes. Things like building footprints or road networks. And then we provide these datasets for free, totally openly to the machine learning computer vision community. And we encourage them to try to develop models that can identify the features that we’ve labeled, and we do this through these challenges that we run with Topcoder, called the SpaceNet challenges.

ABERMAN: And why is that relevant?

WEIR: Well it remains a very open question how well algorithms, computer vision, machine learning, A.I., whatever you want to call it, can do things like identify building footprints in images, extract a road network from an image. And that’s really important because it’s really important to try to accelerate mapping. A lot of the time people will, if they think about mapping, just think of something like google maps, where in the United States, you can just pull out your phone and get a map, and get the best route to pretty much wherever you want.

Subscribe to the What’s Working in Washington podcast on iTunes.

But in a lot of the developing world, this isn’t the case. And then if there’s a sudden change to the environment, this isn’t necessarily the case. Particularly take Puerto Rico after Hurricane Maria, where it was really important for aid organizations and FEMA to have a map to be able to figure out how to best distribute aid, or where damage is, things like this. Now, it took over 5000 volunteer mappers over two months to produce this new map of Puerto Rico, and that’s not a knock on the mappers by any means. It’s a huge effort.

They had to remap almost a million buildings, and something like 30,000 kilometres of roads. And so if we can accelerate that process, so that we can get to a new map faster, after something like a natural disaster hits, or in an underserved area where there’s less commercial benefit to manually creating a map, then it would be fantastic to be able to do that with computer vision.

ABERMAN: I assume this also has enormous national security implications. Battlefield management, trying to identify a threat situation as well.

WEIR: Sure, absolutely. Anything that can be done to accelerate response to changing conditions.

ABERMAN: Now what I find interesting, I want to talk about now, is SpaceNet is a model of what people describe as crowdsourcing. You know, getting the wisdom of large numbers people working a problem, which is very atypical from the way we ordinarily approach government R&D. How is SpaceNet actually using crowdsourcing here?

WEIR: Sure. I’ll talk a little bit more about just how SpaceNet is structured. So what we do is, we generate these datasets where we get imagery from our partners at Maxar, and they also provide manually made labels, drawing actual boundaries around all the buildings, or tracing all the road networks. And then we release these datasets using Amazon Web Services, freely available to pretty much anyone in the world. And so what we do for crowdsourcing, with Topcoder is, we run these coding challenges there, where we will provide a good chunk of this imagery, and a bunch of the labels, to competitors, and say, develop the best algorithm that you can, to identify, say we’re doing building footprints, to identify building footprints from these images.

And then, they will use that algorithm to predict on some imagery that they’ve never seen before. And then, we score how well their algorithms worked. And so, we get usually between 50 and 250 competitors participating in these challenges, and everyone is trying different algorithms to try to solve this problem. Everyone has their own machine learning solution to try to find buildings, and this allows us to figure out which solution works best.

ABERMAN: Let me ask you this: what’s in it for these small teams and entrepreneurs? Why take the time? They don’t all get money, right?

WEIR: No, not all of them do. We give out prize money to the top few competitors, but not to the vast majority. To some degree it’s bragging rights. There are some people who we know do this pretty much for their profession, but the majority of these people are maybe students trying to develop these skills themselves. And this is something that some people can list on their CV as having participated in these competitions.

ABERMAN: So from the government’s perspective, is this talent discovery, or a way to find new vendors?

WEIR: You know in a way, it’s almost neither. What we can do with these challenges, so at the end of the challenge, after the competitors have completed their algorithms, we get the code for the best solutions. And then our small research team, which has far fewer people in it than this crowd source community, can then dig through the best solutions, and figure out what worked well, what didn’t, et cetera. So this is almost a way for us to explore the enormous array of methods that are out there for analyzing imagery, that we couldn’t do alone. We only have four researchers on our team. And for example, I was just at the biggest computer vision conference in the world last week, and there were 1,300 papers in that conference. There’s no way we can keep up with all of that. No matter how hard we try.

ABERMAN: I love this. It seems like a really interesting model. I look forward to hearing about your continued success, and I suspect we’re going to see a lot more of this. Nick Weir, challenge manager for SpaceNet. Thanks for joining us today.

WEIR: Thank you very much for having me, Jonathan.

Copyright © 2024 Federal News Network. All rights reserved. This website is not intended for users located within the European Economic Area.

Related Stories