I love the generic function aspect of this, where there aren't necessarily specific inputs or outputs. This will make code reusable by factors. A simple use being not needing to define both `toRadians` and `toDegrees`, just definite the relationship. But I believe this can be taken to the point of getting backpropagation for free once forward propagation is defined in neural networks.
Not to steal any thunder, but I've been prototyping a language very built on this idea.