“The central proposal of Categorical Deep Learning is that a neural network layer should be viewed as a homomorphism between two algebras for the same endofunctor.”