Do our machines know us better than ourselves? And if they did, could they, in the parlance of the Pentagon, use that knowledge to improve our lethality?
DARPA, the Department of Defense’s blue-sky agency, launched April 29 a program to use artificial intelligence to best match interventions for individuals. It is called “Teaching AI to Leverage Overlooked Residuals,” or TAILOR.
Specifically, DARPA is looking for submissions about how to use AI for “Human Performance Optimization,” or HPO. Crucially, DARPA is looking for alternatives to one-size-fits-all approaches, because universal recommendations based on group averages can work at cross-purposes to individual need.
“This approach frequently (mis)characterizes individual variance as statistical ‘noise,’ ‘residuals,’ or ‘error,’” reads the solicitation. “The resulting interventions (e.g., diet, physical training regimen, brain stimulation) are at best suboptimal and at worst deleterious for each person.”
Diet, physical training, and brain stimulation aren’t the flashiest parts of the military, but they’re fundamental to how everything else operates. DARPA specifically cites an interest from Special Operations Command’s “Close Combat Lethality Task Force” in getting this optimization right, tailoring interventions to individuals and teams into an advantage.
To get there, DARPA is asking proposers to come up with “third wave AI approaches,” to demonstrate contextual reasoning and transfer learning, meaning a solution is capable of adapting to changed circumstances, environments, and can use related knowledge for abstraction. DARPA also wants teams to move away from traditional AI that trains on large data sets, has opaque processes, and are hard to adapt to new contexts. Proposers will be expected to focus on at least on area of human performance, with DARPA aiming to have a full portfolio of research on physical cognitive, and social performance data.
It’s a lot of technical build-up, but the final outcome is relatively straightforward: after given the data and goal for a specific individual or team, the tool needs to evaluate how much that individual or team would benefit from a given intervention (such as a change in diet or exercise or brain stimulation.) The systems will also be evaluated on how well they can explain or accommodate counterfactuals. Phase II is much the same process, but with datasets from other teams, and with evaluation from subject matter experts in government.
Should the research prove fruitful, the military will be one step closer to specific responses for every individual, a personalized tool that’s general issue.