Unverified50% confidenceFactExact time
In AdamW, the parameter update is approximately equal to the first moment divided by (the square root of the second moment plus ε), and the reciprocal of the square root of the second moment acts as a preconditioner, adaptively scaling the gradient for each coordinate.
1
Sources
50%
Confidence
Long-term
Relevance
10/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWhen ε ≈ 0, the second-moment estimate fully dominates and the optimizer reverts to standard Adam.72% similarUnverifiedAdam的一阶矩衰减系数β₁通常默认为0.970% similarUnverifiedAdam的二阶矩估计记录梯度平方的指数移动平均，反映梯度的波动幅度68% similarUnverifiedAdam 优化器由 Kingma 和 Ba 于 2014 年提出，同时维护梯度的一阶矩估计与二阶矩估计，并通过偏差修正项处理训练初期矩估计偏向零的问题65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1013226API
curl https://kongchang.com/api/v1/knowledge/claims/1013226MCP
get_claim(id=1013226)