Sure, I agree the seatbelt example comes from not having enough labeled classes. Feels a little bit more like the realm of semi-supervised learning, but I'd love to learn about interesting approaches in that case too :)