Decoupled weight decay proportional to gamma squared yields stable weight and gradient norms under the steady-state assumption that updates are independent of weights.
The results are mostly expected if we consider the effective learning rateγ eff over time (Fig
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Correction of Decoupled Weight Decay
Decoupled weight decay proportional to gamma squared yields stable weight and gradient norms under the steady-state assumption that updates are independent of weights.