Skip to content

Commit 20a9782

Browse files
Variational ...
1 parent 7c55566 commit 20a9782

File tree

1 file changed

+32
-0
lines changed

1 file changed

+32
-0
lines changed

CNN/second-chapter.md

Lines changed: 32 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,32 @@
1+
# Chapter 2. 불확실성의 언어
2+
3+
## Bayesian modeling
4+
5+
### Likelihood vs Probability
6+
7+
< 특정 x >
8+
불연속 사건 ( 주사위... ) : Likelihood == Probability
9+
연속 사건 ( 신호...? ) : pdf의 y값 == Likelihood || Probability == 0
10+
11+
### Likelihood distribution in Model
12+
13+
$$p(y=d|x,w)=\frac{exp(f^w_d(x))}{\sum_{d'} exp(f^w_{d'}(x))}$$ : Classification tasks
14+
15+
$$p(y=d|x,w)=N(y;f^w(x),t^{-1}I)$$ : Regression....?
16+
17+
#### Predict output for an new input point x*
18+
In Bayesian modeling, inference == integral
19+
$$ p(y^*|x^*,X,Y) = \int p(y^*|x^*,w)p(w|X,Y)dw $$
20+
21+
$$p(Y|X) = \int p(Y|X,w)p(w)dw $$
22+
여기서 계산되지 못할 Inference는 어떻게 처리를 해야하는가....?
23+
24+
## Variational Inference
25+
Function : 입력이 주어지면 이와 매핑된 값을 출력하는 함수 ( ex - $$f(x)=x^2+3$$ )
26+
Functional : 함수들의 집합을 정의역으로 갖는 함수 ( ex - $$J[y] = \int y(x) dx $$ )
27+
-> ( Functional은 결국 loss 함수...? )
28+
Functional의 미분은 y가 다양하기 때문에 '정확히' 계산하는 것은 불가능하다. 이를 Variational Inference를 통해 해결한다 -> 근사화
29+
30+
## Reference
31+
- http://rstudio-pubs-static.s3.amazonaws.com/204928_c2d6c62565b74a4987e935f756badfba.html [가능도(likelihood) vs 확률]
32+
- http://norman3.github.io/prml/docs/chapter10/0

0 commit comments

Comments
 (0)