Hiroki Naganuma

NYU Lecture is really useful https://atcold.github.io/pytorch-Deep-Learning/

6.

Screen Shot 2021-12-02 at 15 53 22

(a).

Describe a scheme to estimate the vector of parameters θ to maximize the log likelihood of the N training examples.

識別モデル:データに対するクラスの条件付き確率(事後確率)p(Ck|x)を直接モデル化するのが識別モデル
生成モデル:p(Ck|x) = p(x|Ck)p(Ck)/p(x) の右辺をモデル化する

Screenshot 2021-12-07 at 4 41 18 PM

(b).

Describe how the scheme should be modified if you are given a prior distribution on the model parameters, p(θ).

Screenshot 2021-12-07 at 11 23 56 AM Screenshot 2021-12-07 at 11 29 02 AM

(c).

Describe the difference between a maximum a posteriori prediction and a full Bayesian prediction. Be as specific as possible.

ベイズ推論の枠組みで、デルタ関数を用いてMAPを記述する

(d).

Screen Shot 2021-12-02 at 15 54 59

最尤法で求めた分散は真の分散を過小評価している

7.

Screen Shot 2021-12-09 at 1 26 58

Screenshot 2021-12-09 at 2 53 07 PM

Screenshot 2021-12-09 at 4 51 37 PM Screenshot 2021-12-09 at 5 00 52 PM