Computer Vision News - October 2021

3 Summary Spleenlab 17 large if-else tree and foresee every potential part in that tree and assess each state at the end of the tree with probabilities. We say our software could run through each part of the tree. We call that determinism . If we have no AI, so-called deep learning, we have a highly statistical system. We can validate our AI, as many people did, but at the end of the day, we have a degree of uncertainty in our system. It’s hard to regulate those systems by a standard that requires determinism because you can’t foresee the last percentage of your system.” U ncertainty is bad . People want to know they are safe in their cars and that drones will not hit other aircraft. But why is there uncertainty in the system? The models are trained on data and validated on data, with millions of parameters in the system that are deterministic, but some methods at inference time, like dropout and pruning make the system statistical. A statistical system always has uncertainty. “Take autonomous driving, for example,” Stefan says. “You have a dataset, you train on the dataset, and then you have a validation dataset that you collect in Germany, and it performs at 99 per cent accuracy . There is still a risk of uncertainty because if you put your validation set to a different domain that's collected in France, with a demographic and geographic domain shift, you may get a 98 per cent accuracy, and you do not really know why. There is no direct if-else connection in a deep learning model.” People are already working on transparent AI , but Stefan says a new standard that explains its deepest behaviors is still many years away, while models are being sold and deployed in the real world now. That is why Spleenlab is focused first and foremost on safety by design. Take the example of an automated drone on a package delivery which encounters an emergency . Its battery has dropped to a critical level. The system has a rest range of 100 meters. There is no known emergency landing point in the area. What does it do? Caroline Essert Landing prediction