SGD:先說清楚衡量什麼SGD rates need a stated metric

IDL115 · CH03 · KP09 · 22/44
非凸看駐點,凸或強凸才可談相應的最優損失差。 / Stationarity and objective gaps are different metrics.

一般平滑非凸Smooth nonconvex setting

一般平滑非凸下,使用加權平均梯度平方作為駐點指標。 / Use an averaged stationarity measure.

GT:=∑k<TηkE∥∇f(wk)∥2∑k<Tηk\mathcal G_T:=\frac{\sum_{k<T}\eta_k\mathbb E\|\nabla f(w_k)\|^2}{\sum_{k<T}\eta_k}

教材式 (32)–(34)Source rates

代入排程得下列量級;非凸時應讀為駐點指標。 / These orders can describe stationarity.

ηk∝1/(k+1): O(1/log⁡T)ηk∝1/k+1: O(log⁡T/T)ηk=η: O(1/(ηT)+η)\begin{aligned}\eta_k\propto1/(k+1)&:\ O(1/\log T)\\\eta_k\propto1/\sqrt{k+1}&:\ O(\log T/\sqrt T)\\\eta_k=\eta&:\ O(1/(\eta T)+\eta)\end{aligned}

強凸情況Strong convexity

在對應條件、排程及輸出選擇下可取得 O(1/T) 最優值差。 / Suitable strong-convexity results improve the rate.

ηk∝1/(μ(k+1)): O(1/T)(35)\eta_k\propto1/(\mu(k+1)):\ O(1/T)\quad\text{(35)}