3 ms·
Can anyone explain: why is the Adam optimizer so unreasonably effective? Is anyone even using a different optimizer anymore?
by amelius 4y ago
Can anyone explain: why is the Adam optimizer so unreasonably effective? Is anyone even using a different optimizer anymore?
- jonbaer 4y agoWhat is the current SOTA? Demon Adam? https://arxiv.org/pdf/1910.04952v4.pdf https://arxiv.org/pdf/1910.04952v4.pdf
- NavinF 4y agoIt’s really hard to compare optimizers. Common architectures and default hyperparameters were discovered alongside Adam so you’d have to redo a bunch of sweeps if you wanted a “fair” comparison. In practice this doesn’t really matter and everyone just uses Adam. If you had infinite compute, you’d try every combo and select the one with the best results.
- Yenrabbit 4y agoA good run-down of the different algorithms: https://ruder.io/optimizing-gradient-descent/ https://ruder.io/optimizing-gradient-descent/ I think people favour adaptive learning rate options like Adam in practice since they generally do seem to perform well, and are often less sensitive to initial conditions and the exact hyper-parameters used. There will always be people who like to test N optimizers with parameter sweeps to squeeze a tiny bit of extra performance out, but for the rest of us the default Adam or AdamW options and good, unobjectionable choices :)
- yobbo 4y agoAdam functions as a low-pass filter (and/or "compressor") on the gradient. It filters out "noise", which is "wild" during start of training. This is basically what all "optimizers" achieve in various ways, including momentum. Adam uses several times more memory, and is slower, than momentum or just SGD. That's reason to not use if not needed.
- locuscoeruleus 4y agoAdam was very effective when it got introduced so it was widely adopted. Since then only models that work well with Adam have made it from the idea stage to actually working. I think there's reason to believe we have over fit our model architectures to our loss functions and optimizers.