Notice
Recent Posts
Recent Comments
Link
«   2025/04   »
1 2 3 4 5
6 7 8 9 10 11 12
13 14 15 16 17 18 19
20 21 22 23 24 25 26
27 28 29 30
Tags
more
Archives
Today
Total
관리 메뉴

statduck

Exponential Family 본문

잡담

Exponential Family

statduck 2023. 6. 20. 00:37

우리가 흔히 알고있는 분포들(이항분포, 다항분포, 정규분포)은 주로 지수족(Exponential Family)에 속한다.

지수족은 몇가지 특성들을 만족시키는 분포들의 집합이라고 생각하면 된다.

정의로는 다음과 같은 확률함수의 모양을 가지면 지수족이다.

p(x|η)=1Z(η)h(x)exp[ηTT(x)]=h(x)exp[ηTT(x)A(η)]

  • ηRK with fixed support over xDRD
  • η: canonical parameters, T(x): suffcient statistics
  • A(η)=logZ(η): log partition function ( A is a convex function over the convex set Ω)

다음의 세가지 이유로 지수족은 유용하다고 여겨진다.

  • Log partition function can generate moments by its derivatives
  • Covaraicne of the sufficient statistics are the same as Fisher Information Matrix
  • The statistics of moments are easily derived from T(x)(even MLE)

첫번째 불릿 - Log partition function can generate moments by its derivatives

부터 살펴보자. 다음 두가지가 매우 중요하다. ηA(η)=E[T(x)],2ηA(η)=Cov[T(x)] 이 중 첫번째에 대한 증명이 아래와 같다.

결국 핵심은 미분으로 모멘트를 생성할 수 있다는 것이다.

 

두번째 불릿 - Covaraicne of the sufficient statistics are the same as Fisher Information Matrix(FIM)

몇개의 조건(regulartiy conditions) 하에서 Fisher Information은 다음과 같이 계산된다.

F(η)=Ep(xη)[2ηlogp(x|η)]

F(η)=Ep(x|η)[2η(ηTT(x)A(η))]=2ηA(η)=Cov[T(x)]

즉, 로그 분할 함수(log partition function)의 헤시안 매트릭스가 되는데, 이는 충분통계랑 T의 공분산과 동일하다는 것이다.

Cramer-Rao Lower Bound 등을 계산할 때 FIM이 이용되는데 지수족에 있는 분포들이면 손쉽게 계산된다.

 

세번째 불릿 -The statistics of moments are easily derived from T(x)(even MLE)

The likelihood of an exponential family has the form

p(D|η)[ΠNn=1h(xn)]exp(ηT[Nn=1T(xn)]NA(η))exp[ηTT(D)NA(η)]

  • T(D)=[Nn=1T1(xn),...,Nn=1TK(xn)]
  • 이 때 해당 꼴에서 로그를 씌운 로그 우도함수(log-likelihood function)의 미분값이 0이 되면 이 때 충분통계량의 추정평균과 실제 평균이 같아진다.

logp(D|η)=ηTT(D)NA(η)+const

ηlogp(D|η)=ηηTT(D)NηA(η)=T(D)NE[T(x)]

N=1(single data case)에 대해서는 다음과 같다.

ηlogp(x|η)=T(x)E[T(x)]

이 때 좌변의 gradient를 0으로 조절하면 (로그 우도함수의 최대화) 다음과 같으며 이를 moment matching 이라 한다.

E[T(x)]=1NNn=1T(xn)

 

 

Probabilistic Machine Learning: Advanced Topics. probml.github.io. (n.d.). https://probml.github.io/pml-book/book2.html

'잡담' 카테고리의 다른 글

게임에 적용되는 ML/DL  (0) 2022.09.27