This paper is all about trying a bunch of different changes to the training setup to see what affects the power law exponent over dataset size. Here are some of the answers:
- encoder-decoder size asymmetry: exponent not affected, but effective model capacity affected
- architecture (LSTM vs. Transformer): exponent not affected, but effective model capacity affected
- dataset quality (filtered vs. not): exponent and effective model capacity not effected, losses on smaller datasets affected
- dataset source (ParaCrawl vs. in-house dataset): exponent not affected
- adding independent noise: exponent not affected, but effective model capacity affected
Here are some other things to test that I thought of while I read this:
- compare scaling with respect to language pairs (the architecture experiments saw
\(p=0.28\)
and\(p=0.25\)
for en -> de and zh -> en respectively. Is that difference significant?)