I'm trying to understand the basics of quantization in Neural networks. Quantization tries to convert a neural network that uses floating point arithmetic to one that uses a lower precision integer arithmetic (mostly during inference). While searching around about it, I ran into the Gemmlow documentation that gives a good introduction to the concept.
I understood most of what they mention in that documentation except the first sentence of the Quantization as an affine map section that mentions that in order for the arithmetic to map from floating point to integer directly, only affine functions might work.
Now I understand the basics of affine function a bit, but I couldn't really understand why only affine functions make sense. Could someone explain if there's any property of affine functions that might be missing in other types of functions which make this statement true.
