Computational complexity of neural networks (2022) (lunalux.io)

This is simply wrong. Backprop has the same asymptotic time complexity as forward.

"If we once again assume that there are the same number of neurons in each layer, and that the number of layers equal the number of neurons in each layer we find:"

These are terrible assumptions.

Why not compute the runtime as the product of the actual sizes? Then the comparison will also make more sense.