Is AI/ML data too unique for the DoD to stick into one centralized data repository?

Not only is the data dirty, but it’s unique. The satellite derived data that NGA [National Geo-spatial Intelligence Agency] is using is, while it might still be imagery that we take from mid- to high-altitude UAS platforms, it’s completely different — even though it might still be imagery. The metadata is different, the was we tag it for usage in machine learning applications is different. It’s really hard. In the DoD world we’ve probably all heard about these offices and organizations that are going to be stood up to handle and build one central data repository. It’s my belief that will not work. The data is too unique. You cannot create one place where we can go get data for all machine learning applications.

That was Rob Albritton on The Convergence podcast, “AI across the enterprise.” I tend to agree with Albritton on that point, but that doesn’t mean that there cannot be several data repository efforts for domains where it makes sense. We shouldn’t be aggregating data assuming that general AI will come along. Instead, optimize for speed and creativity by allowing each service to have a couple different repositories, each will be better suited to enabling a wide array of narrow AI application.

Based on the narrowness of AI for the foreseeable future, I think there’s a lot of room for deception as an effective countermeasure. Present an AI application with something it didn’t expect, or ask it to answer a question it wasn’t trained for, and it will stop being so effective. Here is Albritton on the subject:

It’s all that far-fetched to think that in 5 to 10 years, our soldiers could experience an encounter with fully autonomous squads of weapon systems. Think robotic infantry. If we were every to go to battle with a near peer like Russia in Eastern Europe, those kind of adversaries have no qualms using fully autonomous weapon systems, where our ethics and our morals don’t allow us to do that. How do we combat that? … I think the answer is looking at things like counter AI, and how do we jam systems or deceive fully autonomous systems? The reality is that we’re not going to be willing to build fully autonomous systems to compete with them.