If you are an AGI/deep learning worrier, here is some good news: the seems are starting to show.
By some strange coincidences of cavities and countermeasures, my family have taken to watching videos to combat boredom while brushing two minutes. After a series of experimentation, we discovered the best way to watch a tiny screen is to watch its reflection on the vanity mirrors–the phone or iPad facing away from us. The sound quality is also admirable having bounced off of the same surface.
I discover that I’m able to unlock the iPad using FaceID in reverse. This is iPadOS 13.1.3 circa q4 2019. This seems to be an artifact of ConvNet training that renders the network insensitive to reflections. (Most naively if you imagine the trainer showed the network each image but each time modified the image by adding splotches or shifting or rotating or reflecting) This was popular in the 2010’s because it builds a more robust network that can more reliably detect objects.
While there might be a chance that it is necessary to unlock in the mirror to defuse a bomb in a Sci-Fi movie, the more likely opinion comes from security purists: you can fool someone to look into a mirror and unlock their iPhone or iPad. The most likely thing that happened was that reflection locking was introduced inadvertently applying sota DL algorithms.
This is a very exciting field with a lot of problems to solve! Let’s see if this bug/feature is fixed/justified respectively by the time you see this.