題目(145):機器學習中的優化問題,哪些是凸優化問題,哪些是非凸優化問題?請各舉一個例子。
凸優化問題
邏輯迴歸
Li(θ)=log(1+exp(−yiθTxi))
損失函數推導
logistic regression model:
log1−pp=θTx⇒p=1+exp(θTx)exp(θTx)
maxMLE≃−minlogMLE:=minL(x,y;θ)
L=−(ylogp+(1−y)log(1−p))=−ylog1+exp(−θTx)1−(1−y)log1+exp(θTx)1=ylog(1+exp(−θTx))+(1−y)log(1+exp(θTx))=log(1+exp(−θTx⋅y)),
where Y∈{0,1} and p=P(Y=1∣X=x).
其它例子:SVM, linear regression
非凸優化問題
PCA
VVTminL(V)=∥X−VTVX∥F2
(minimise the reconstruction error)
Formulation from the perspective of maximising the variance
驗證該目標爲非凸問題:檢查定義
If V∗ is the minimum, then −V∗ is also the minimum as L(V∗)=L(−V∗).
L(21V∗+21(−V∗))=L(0)=∥X∥F2>∥X−V∗TV∗X∥F2=21L(V∗)+21L(−V∗)
求解: SVD
其它例子:low-rank model (e.g. matrix decomposition), deep neural network
參考文獻:
- 《百面機器學習》