SGD with momentum can match Adam's performance in language modeling when trained with small batches and careful tuning, a result that contradicts several popular explanations for the optimizer gap.
As before, the gap decreases the longer we train, and SGD can eventually outperform Adam
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Is your batch size the problem? Revisiting the Adam-SGD gap in language modeling
SGD with momentum can match Adam's performance in language modeling when trained with small batches and careful tuning, a result that contradicts several popular explanations for the optimizer gap.