The vast majority are explicitly not deterministic. Starting with random initialization, and not ending there (explicitly introducing noise, using different (an

The vast majority are explicitly not deterministic. Starting with random initialization, and not ending there (explicitly introducing noise, using different (and very low) levels of accuracy, random choices for updates, ...) It is often explicitly mentioned in articles, currently mostly for reinforcement learning (or it might be that that's just the vast majority of papers I see these days), that, for example "this network converges about 3 times out of every 10 attempts". Often with graphs leaving a selection of the failed attempts. This is very different from the explicit lying you find in psychology papers and theories.