做个记录,
1,Isx,y - Ix,y = I(s,y)|x
讨论:
1) x~粗粒度时,求证分段I(s,y)|x~最大 => I(s,y)|x最大,(反之不一定成立。)
2) 每次取一个x~的sy组,多次sgd效果相同,因为L=-<i(s,y)|x>,是多次x采样的均值。
2, 若s|xy=s|x:
a, 则:I(s,y)|x = 0
b, 若I(x,y)|s = 0, 则 Is,y = Ix,y.
-Isn,y + Csn|y - Cs0 + Sum_l[Cs_|sl]
=-Iin,y + Sum_l[<log(Pil|s_)>] - Sum_l[<log(Qi_|sl)>] - <log(Rin)>
其中:
<log(Pil|s_)> = <nlog[vil(s_)]>/2 - log(2pi)/2 - 1/2
-<log(Qi_|sl)> = -<nln[vi_(sl)]>/2 + log(2pi)/2 + (i_ - mi_)**2/(vi_)**2/2
-<log(Rin)> = in**2/2 + log(2pi)/2
网友评论