일 | 월 | 화 | 수 | 목 | 금 | 토 |
---|---|---|---|---|---|---|
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 |
- ddpm
- Stat110
- animation retargeting
- 딥러닝
- ue5.4
- BERT
- GAN
- dl
- RNN
- 생성모델
- multimodal
- 모션매칭
- 언리얼엔진
- Few-shot generation
- WinAPI
- UE5
- CNN
- 오블완
- deep learning
- motion matching
- 디퓨전모델
- Generative Model
- WBP
- Font Generation
- NLP
- Diffusion
- cv
- Unreal Engine
- userwidget
- 폰트생성
- Today
- Total
Deeper Learning
[Statistics 110] Lecture 20: Multinomial and Cauchy 본문
[Statistics 110] Lecture 20: Multinomial and Cauchy
Dlaiml 2022. 12. 25. 17:302-D LOTUS
Z1,Z2가 i.i.d 표준정규분포를 따를 때, E(|Z1−Z2|)를 구하라
X,Y가 서로 독립이고 정규분포를 따를 때, X+Y∼N(μ1+μ2,σ21+σ22).
Z1−Z2는 N(0,2)를 따르기 때문에 Z=x−μσ에서 N(0,2)=√2Z.
√2E(|Z|)=√2∫−inf
Multinomial Distribution(다항분포)
\vec X\sim Mult(n,\vec p), \vec p=(p_1,...,p_k), p_j \le 0, \sum p_j = 1.
여러 확률변수가 포함된 joint distribution
본 코스에서 처음 배우는 다변량분포
이항분포의 일반화 버전
Story
k개의 카테고리로 나눌 수 있는 object가 n개 존재. p_j는 object가 카테고리 j에 속할 확률.
X_j는 카테고리 j에 속하는 object의 개수
Joint PMF
P(X_1=n_1, ...,X_k=n_k) = \frac{n!}{n_1!n_2!...n_k!}p_1^{n_1}p_2^{n_2}...p_k^{n^k}, (n_1+...+n_k = n)
marginal distribution
\vec X\sim Mult_k(n,\vec p).
X_j의 marginal distribution 구하기
카테고리 j에 속하는 시행을 성공, 아닐 경우를 실패라고 생각해보자. 바로 이항분포와 같은 형태다.
X_j \sim Bin(n, p_j).
Conditional joint distribution
\vec X \sim Mult_k(n,\vec p), given \space X_1=n_1.
(X_2,...,X_k) \sim Mult_{k-1}(n-n_1, (p'_2, p'_3,...,p'_k))
단순하게 p_1를 제외하고 식을 쓰면 p_j의 모든 합이 1이 되지 않기 때문에 명백하게 오류가 있음을 알 수 있음.
p'_2는 P(카테고리 2에 속함 | 카테고리 1에 속하지 않음)이기 때문에 조건부확률 식을 이용하여 풀면
p'_2 = \frac{p_2}{1-p_1} = \frac{p_2}{p_2+...+p_k}\\p'_j = \frac{p_j}{p_2+...+p_k}
Lumping Property
\vec X = (X_1,X_2,...,X_{10}) \sim Mult(n,(p_1,...,p_{10})).
어떤 나라에 10개의 정당이 있고 n명의 국민들은 모두 어느 정당에 속해야 한다라는 이야기로 표현
X_j는 j정당에 속한 사람의 수, p_j는 j정당에 속할 확률
주력 정당인 1, 2정당을 제외한 3~10 정당을 “비주류 정당”이라고 이름 붙여 묶고 다시 식을 써보면
\vec Y = (X_1, X_2, X_3+...+X_{10}) \sim Mult(n,(p_1,p_2,p_3+...+p_{10}))
Cauchy Distribution(코시분포)
Cauchy Interview Problem
코시분포는 X,Y가 i.i.d 표준정규분포 N(0,1)일 때 \frac XY의 분포
기댓값이 발산
joint CDF (\Phi는 표준정규분포의 cdf)
P(\frac{X}{Y} \le t) = P(\frac{X}{|Y|} \le t) = P(X \le t|Y|)\\ = \frac{1}{\sqrt{2\pi}} \int_{-\inf}^{\inf}e^{-y^2/2}\int_{-\inf}^{t|y|}e^{-x^2/2}dxdy \\= \frac{1}{\sqrt{2\pi}} \int_{-\inf}^{\inf}e^{-y^2/2}\Phi(t|y|) dy\\ =\sqrt{\frac{2}{\pi}} \int_{0}^{\inf}e^{-y^2/2}\Phi(ty) dy,(even \space function)
이제 적분을 하면 되는데 문제는 \Phi가 적분이 불가능한 intractable 함수로
모든 구간에서 연속이고 미분가능한 성질을 지니면(미분적분학 내용) 미분과 적분을 바꿔도 문제가 없는 well behaved 함수
PDF부터 구하면
F'(t) = f(t) = \sqrt{\frac{2}{\pi}} \int_{0}^{\inf}ye^{-y^2/2}\frac {1}{\sqrt{2\pi}}e^{-t^2y^2/2} dy \\ = \frac{1}{\pi} \int_0^{\inf}ye^{-(1+t^2)y^2/2}dy \\u = (1+t^2)y^2/2, \frac{du}{dy} = y(1+t^2)\\du = y(1+t^2)dy \\ \frac{1}{\pi} \int_0^{\inf}ye^{-(1+t^2)y^2/2}d = \frac{1}{\pi} \int_0^{\inf}\frac1{1+t^2}e^{-u}du\\ = \frac1{\pi(1+t^2)}
CDF는 PDF를 적분한 arctan이 포함된 수식
LOTP 활용
P(X \le t|Y|) = \int P(X \le t|Y||Y=y)f_Y(y) dy
독립이기 때문에 P(X \le (t|Y|) | Y=y)에서 조건 y를 대입 가능 (독립이 아니면 조건식을 없앨 수 없음)
\int P(X \le t|Y||Y=y)f_Y(y) dy = \int P(X \le t|y|)f_Y(y) dy = \int \Phi(t|y|) f_Y(y) dy
Reference
[0] https://www.youtube.com/playlist?list=PL2SOU6wwxB0uwwH80KTQ6ht66KWxbzTIo
Statistics 110: Probability
Statistics 110 (Probability) has been taught at Harvard University by Joe Blitzstein (Professor of the Practice in Statistics, Harvard University) each year ...
www.youtube.com
'Statistics & Math > Statistics 110: Probability' 카테고리의 다른 글
[Statistics 110] Lecture 22: Transformations and Convolutions (0) | 2022.12.25 |
---|---|
[Statistics 110] Lecture 21: Covariance and Correlation (0) | 2022.12.25 |
[Statistics 110] Lecture 19: Joint, Conditional, and Marginal Distributions (0) | 2022.12.24 |
[Statistics 110] Lecture 18: MGFs Continued (0) | 2022.12.24 |
[Statistics 110] Lecture 17: Moment Generating Functions (0) | 2022.12.24 |