In amplification, we need to decompose a task into parts such that we can evaluate performance on each part separately. That’s what lets us train the parts separately and then combine them, which is what lets the whole thing get off the ground.

In a CNN, without doing some extra work, we don’t have any decomposition into semantically meaningful parts whose performance can be evaluated separately (instead we just evaluate the performance of the whole thing). So it’s not useful for amplification.