一階與二階:資訊與成本First and second order: information and cost
一階看坡度,二階再看坡度如何改變。 / First order uses slopes; second order also uses curvature.
一階方法First-order methods
GD、SGD、Adam 用梯度決定更新。 / GD, SGD and Adam use first derivatives.
二階方法Second-order methods
Newton 法解線性系統;通常不直接計算矩陣反矩陣。 / Newton methods solve a linear system.
規模差異Scaling
d 個參數:梯度有 d 個元素;稠密 Hessian 有 d² 個元素。 / A dense Hessian has d² entries.