Calibrating AI For Saftey

Greetings from the Australian National University, where I attended the weekly AI, ML and Friends seminar. Jiawei Liu was speaking on "Uncertainty Calibration for Deep Neural Networks". I didn't understand the equations, but the idea, as I understand it, is to have a measure of how well AI is doing.

What got my attention was an example of recognising a speedboat. From the original photo it was clear to a human what this was. But the black and white outline, as used by some algorithms, looked to me like a USV (Uncrewed  Surface Vessel). These are being used, right now, in the Black and Read Seas, laden with explosives, to attack ships. The crews of warships have to maintain a lookout 24 hours a day, for weeks, trying to spot drone attacks. The USVs are modified speedboats so it is difficult to tell them from fishing boats. This type of AI analysis can help ensure mistakes are not made which could result in the loss of a ship, or the sinking of an innocent fishing boat.