WEBVTT
1
00:00:00.000 --> 00:00:08.477
[MUSIC]
[淺談貝氏統計]
2
00:00:08.477 --> 00:00:12.890
In this lecture we'll talk
about Bayesian statistics.
這一講來談談貝氏統計
3
00:00:12.890 --> 00:00:16.190
In a previous lecture we talked about
likelihoods and expressing relative
前一講介紹似然性,以及如何表達證據支持兩種假設的相對程度
4
00:00:16.190 --> 00:00:19.410
evidence for one hypothesis compared
to an alternative hypothesis.
前一講介紹似然性,以及如何表達證據支持兩種假設的相對程度
5
00:00:19.410 --> 00:00:24.810
And we've used evidence in this sense just
based on the data that we have at hand.
這讓我們能用手上的資料做為證據
6
00:00:24.810 --> 00:00:27.960
But we'll see that very often you have
some prior belief that you might want to
貝氏統計能結合證據與根據假設的事前信念(prior),計算推論的可靠性
7
00:00:27.960 --> 00:00:31.460
incorporate with the evidence and
Bayesian statistics allows you to do this.
貝氏統計能結合證據與根據假設的事前信念(prior),計算推論的可靠性
8
00:00:32.560 --> 00:00:34.490
Now let's say that you
flip a coin three times.
假想一枚硬幣丟了三次
9
00:00:35.680 --> 00:00:38.860
The coin comes up heads every single time.
每一次都是正面朝上的結果
10
00:00:38.860 --> 00:00:41.420
Now do you think that this
is a fair coin or not?
根據這個結果能判斷這是枚公正硬幣嗎?
11
00:00:42.670 --> 00:00:46.440
The data that you have at hand shows that
this coin came up heads three times.
因為眼前的資料是三次正面朝上
12
00:00:46.440 --> 00:00:52.380
So if you would be a newborn baby
then you don't have any prior beliefs and
對一位全無任何事前信念的新生兒來說
13
00:00:52.380 --> 00:00:54.520
all you have is the data
that you just observed.
只有眼前看到的資料
14
00:00:54.520 --> 00:00:59.810
So for a newborn
it comes up heads every single time.
新生兒會認為這枚硬幣總會擲出正面朝上
15
00:00:59.810 --> 00:01:02.030
This is just what this coin does.
正如現在看到的事實
16
00:01:02.030 --> 00:01:05.100
But you have prior beliefs
that most coins are fair.
對於會看這部影片的大人來說,都有多數硬幣是公正的信念
17
00:01:05.100 --> 00:01:07.230
You might think
你也許認為就算現在是這種結果
18
00:01:07.230 --> 00:01:11.030
Three heads in a row
that's a perfectly likely observation.
也只是剛好看到三枚正面朝上而已
19
00:01:11.030 --> 00:01:15.250
I'm not changing my prior beliefs
that this is a fair coin yet.
你不會因此改變這是一枚公正硬幣的信念
20
00:01:15.250 --> 00:01:16.490
You would want to see more evidence.
你會想要再做更多試驗
21
00:01:18.670 --> 00:01:24.010
Now
data is possible in this Bayesian sense.
貝氏統計能結合個人信念與事實證據
22
00:01:24.010 --> 00:01:27.260
It's not possible when you
just calculate P-values.
這是p值做不到的功能
23
00:01:27.260 --> 00:01:31.136
Remember that the P-value expresses
the probability of the data or
請記住p值表達以虛無假設為真的前提,不論手上資料是否極端,確實存在的程度
24
00:01:31.136 --> 00:01:34.678
more extreme data assuming that
the null hypothesis is true.
請記住p值表達以虛無假設為真的前提,不論手上資料是否極端,確實存在的程度
25
00:01:34.678 --> 00:01:38.544
And some people say
actually not what you want to know.
有人認為p值不能代表研究者想要得到的最終解答
26
00:01:38.544 --> 00:01:42.727
What you want to know is the probability
that the null hypothesis is true
研究者要追求的是根據手上資料,虛無假設確實存在的程度
27
00:01:42.727 --> 00:01:45.092
given some data that you have collected.
研究者要追求的是根據手上資料,虛無假設確實存在的程度
28
00:01:45.092 --> 00:01:47.077
And this is a posterior probability.
這就是事後機率(posterior probability)
29
00:01:47.077 --> 00:01:50.699
Given that you've collected some data
maybe some prior beliefs that you have.
根據手上的資料,加上已存在研究者腦中的信念
30
00:01:50.699 --> 00:01:53.532
What is now the probability
that the null hypothesis or
能不能獲得虛無假設或對立假設為真的機率?
31
00:01:53.532 --> 00:01:56.240
the alternative hypothesis is true?
能不能獲得虛無假設或對立假設為真的機率?
32
00:01:56.240 --> 00:01:59.090
So Bayesian statistics allows
you to express this probability.
貝氏統計正是用來計算事後機率
33
00:02:01.090 --> 00:02:02.630
It works quite straightforward.
貝氏統計的原則很簡單
34
00:02:02.630 --> 00:02:05.890
You have some prior belief
you have some data.
把腦中的事前信念加上手上的資料
35
00:02:05.890 --> 00:02:08.110
You combine these into a posterior belief.
就能獲得目前證據的事後機率
36
00:02:10.530 --> 00:02:14.935
We can calculate posterior odds that
the alternative hypothesis is true given
根據手上資料計算的事後機率比(posterior odds)
可以比較對立假設為真與虛無假設為真的程度
37
00:02:14.935 --> 00:02:15.535
the data
根據手上資料計算的事後機率比(posterior odds)
可以比較對立假設為真與虛無假設為真的程度
38
00:02:15.535 --> 00:02:19.884
compared to the probability that the null
hypothesis is true given the data.
根據手上資料計算的事後機率比(posterior odds)
可以比較對立假設為真與虛無假設為真的程度
39
00:02:19.884 --> 00:02:25.499
This is a very useful way to split apart
these different aspects of the formula.
對照公式的每個部分與貝氏原則可以看到
40
00:02:25.499 --> 00:02:31.090
So the posterior equals
the likelihood ratio times the prior.
事後機率就是似然性比與事前信念的乘積
41
00:02:31.090 --> 00:02:33.224
So
this likelihood ratio
公式裡的似然性比
42
00:02:33.224 --> 00:02:35.138
that we discussed in a previous lecture
就是前一講介紹的重點
43
00:02:35.138 --> 00:02:38.818
is an essential part of the probability
calculations in Bayesian statistics.
可以看到似然性比是計算事後機率的重要成份
44
00:02:38.818 --> 00:02:42.672
But it's also combined with this prior
so how can we do this?
不過要結合事前機率,那麼要怎麼知道事前機率呢?
45
00:02:42.672 --> 00:02:47.050
We have to come up with
a prior distribution.
在此介紹一種事前機率分配
46
00:02:47.050 --> 00:02:51.670
In the case of binomial probabilities
a beta distribution is used.
所謂的beta分配是用在資料有二項性質的狀況
47
00:02:51.670 --> 00:02:55.450
The beta prior is determined by two
parameters which are referred to as
beta事前分配由alpha與beta兩種參數決定
48
00:02:55.450 --> 00:02:56.200
alpha and beta.
beta事前分配由alpha與beta兩種參數決定
49
00:02:57.380 --> 00:03:01.580
Now this can get a little confusing
because we already talked about alphas and
在這裡你也許會和之前提過代表型一與型二錯誤率的alpha與beta搞混
50
00:03:01.580 --> 00:03:04.079
betas as type one errors and
type two errors.
在這裡你也許會和之前提過代表型一與型二錯誤率的alpha與beta搞混
51
00:03:05.700 --> 00:03:09.840
Regrettably
are not very creative when it comes to
這是因為統計學家使用希臘字母做統計符號的創意不夠所致
52
00:03:09.840 --> 00:03:13.280
thinking up greek letters that they
want to use for their statistics.
這是因為統計學家使用希臘字母做統計符號的創意不夠所致
53
00:03:13.280 --> 00:03:15.530
So there are some double
use in the literature
有時一種符號在同一篇論文裡有很多意思
54
00:03:15.530 --> 00:03:17.320
this is one of these examples.
beta分配就是一個例子
55
00:03:17.320 --> 00:03:19.820
The beta prior is determined by alpha and
beta.
這裡的alpha與beta只是決定beta分配的性質
56
00:03:19.820 --> 00:03:23.010
These are just two numbers that have
nothing to do with error rates whatsoever.
與任何形式的錯誤率全無關係
57
00:03:24.900 --> 00:03:28.670
Let's take a look a different
versions of prior distributions.
來看看幾種不同的beta事前分配
58
00:03:28.670 --> 00:03:32.761
If we set the alpha and
the beta in the beta probability to 1
這道beta分配的alpha與beta都是1
59
00:03:32.761 --> 00:03:35.659
we see that we got
the perfectly flat line.
呈現出一條完美的水平線
60
00:03:35.659 --> 00:03:40.259
This means that every value
of theta which is plotted on
也就是說得到範圍內任一統計值(theta)的機率都是相等的
61
00:03:40.259 --> 00:03:43.659
the horizontal axis is equally likely.
也就是說得到範圍內任一統計值(theta)的機率都是相等的
62
00:03:43.659 --> 00:03:46.100
So we don't have any expectations and
這種分配符合不預期特定結果的信念
63
00:03:46.100 --> 00:03:49.160
we say anything that
might happen is possible.
可以說任何結果都有可能發生
64
00:03:50.580 --> 00:03:52.570
Now in the case of a coin flip
在投擲硬幣的例子裡
65
00:03:52.570 --> 00:03:55.920
you might not be convinced that
this is what's most likely.
就是代表不會有經常觀察到某種結果的信念
66
00:03:55.920 --> 00:04:00.110
In the case of flipping a coin
you can have a lot of different beliefs.
當然對於硬幣是否公正的信念可以因人而異
67
00:04:01.280 --> 00:04:06.000
One of them is the belief that this
is a coin that always comes up heads.
其中一種信念是想信這枚硬幣永遠擲出正面朝上
68
00:04:06.000 --> 00:04:09.990
If you have a very strong belief that this
is the case
若你相信如此,這道beta分配完美體現你的想法
69
00:04:09.990 --> 00:04:12.665
You think that it will
be heads almost always
雖然少數情況會出現不同的結果(theta接近0),多數情況都會預期正面朝上(theta接近1)
70
00:04:12.665 --> 00:04:17.341
although you are willing to give at least
some probability of different values.
雖然少數情況會出現不同的結果(theta接近0),多數情況都會預期正面朝上(theta接近1)
71
00:04:17.341 --> 00:04:21.183
You might also think that this
is a perfectly fair coin
你也可能根據多數硬幣是公正的,而相信眼前這一枚也是
72
00:04:21.183 --> 00:04:23.320
most coins are fair after all.
你也可能根據多數硬幣是公正的,而相信眼前這一枚也是
73
00:04:23.320 --> 00:04:27.670
So you think that this coin should
come up heads 50% of the time.
你認為每次擲出正面的機率是50%
74
00:04:27.670 --> 00:04:29.300
You have a very strong belief in this.
把這樣的信念轉換成beta分配
75
00:04:30.320 --> 00:04:33.210
And it's illustrated in this graph
where there's a very high peak
分配就是theta等於0.5之處有一個高峰
76
00:04:33.210 --> 00:04:34.810
at 0.5 on the theta.
分配就是theta等於0.5之處有一個高峰
77
00:04:35.940 --> 00:04:39.450
But you're also willing to allow for
some variation.
當然你也可以認為存在其它可能的結果
78
00:04:39.450 --> 00:04:42.580
It might be that the coin
is not perfectly fair.
因為一些原因,眼前這枚硬幣並不公正
79
00:04:42.580 --> 00:04:45.080
It might have been used and
it might be an old coin and
也許被反覆使用,也許太舊了
80
00:04:45.080 --> 00:04:48.249
might come up heads a little bit
more often or a little less often.
使得擲出正面的次數多一些
81
00:04:51.710 --> 00:04:55.030
Now let's combine our prior belief
whatever it is
來看看怎樣把這些代表事前信念的beta分配
82
00:04:55.030 --> 00:04:57.230
with the data that we've observed.
與觀察獲得的資料結合在一起吧
83
00:04:57.230 --> 00:05:01.810
So we have beta prior distribution
and we have the likelihood function
現在有beta分配代表事前機率
手上資料可轉換成似然性函數
84
00:05:01.810 --> 00:05:06.030
and we can quite easily combine
these into a posterior distribution.
結合兩者就能立刻獲得事後機率分配
85
00:05:08.240 --> 00:05:12.840
The posterior is also a beta distribution
with an alpha and a beta value.
事後機率分配也是一道有alpha與beta值的beta分配
86
00:05:12.840 --> 00:05:16.650
And the alpha and
the beta are determined simply by
要計算事後分配的alpha和beta,只要把事前的參數加上似然性函數估計的參數,再減去1就行了
87
00:05:16.650 --> 00:05:21.640
adding the alpha of the prior to the alpha
of the likelihood function minus 1.
要計算事後分配的alpha和beta,只要把事前的參數加上似然性函數估計的參數,再減去1就行了
88
00:05:21.640 --> 00:05:23.470
The same is true for the beta.
要計算事後分配的alpha和beta,只要把事前的參數加上似然性函數估計的參數,再減去1就行了
89
00:05:23.470 --> 00:05:26.190
The beta of the prior
distribution added to the beta
要計算事後分配的alpha和beta,只要把事前的參數加上似然性函數估計的參數,再減去1就行了
90
00:05:26.190 --> 00:05:27.964
of the likelihood distribution minus 1.
要計算事後分配的alpha和beta,只要把事前的參數加上似然性函數估計的參數,再減去1就行了
91
00:05:29.590 --> 00:05:30.710
Let's see how this works.
來看實際的例子
92
00:05:32.110 --> 00:05:36.760
In this case we have a prior distribution
that's uninformative
這道事前分配是無資訊的(uninfomative),是最容易理解的情況
93
00:05:37.920 --> 00:05:40.972
So this is a prior before
we collect some data.
代表未收集資料前認定任何結果發生的機率都相等
94
00:05:40.972 --> 00:05:44.600
And now we collect some data and
we have a likelihood function.
收集資料後有了這道似然性函數(藍色虛線)
95
00:05:44.600 --> 00:05:47.510
In this case we flip the coin ten times.
這道函數表示投了硬幣十次
96
00:05:47.510 --> 00:05:50.810
Six out of the ten times
we observed heads.
其中六次看到正面朝上
97
00:05:50.810 --> 00:05:53.550
So that's expressed by this
likelihood function and
所以看到高峰出現在theta等於6
98
00:05:53.550 --> 00:05:57.160
now we will combine these two functions
into a posterior distribution.
接著把兩道分配結合在一起,就會看到事後分配的模樣
99
00:05:59.590 --> 00:06:04.600
You now see a black line that falls
exactly on top of the likelihood function.
事後分配的黑線與似然性函數的曲線完美疊合
100
00:06:04.600 --> 00:06:07.730
This is what's meant with
an uninformative prior.
這是因為事前分配是無資訊的
101
00:06:07.730 --> 00:06:10.423
Everything was equally likely beforehand
and
事前認為任何結果都有可能
102
00:06:10.423 --> 00:06:13.123
it doesn't influence our
judgments in any way.
這樣的信念不影響對結果的判斷
103
00:06:13.123 --> 00:06:16.214
We just believe whatever we
have observed exactly the same.
所以事後分配與實際資料一致
104
00:06:16.214 --> 00:06:19.800
So the prior does not influence
the posterior distribution in any way.
因此這個例子事前信念毫不影響事後分配
105
00:06:21.180 --> 00:06:23.630
Let's take another example
where this is the case
接著來看另一個事前信念確實影響事後分配的例子
106
00:06:23.630 --> 00:06:26.680
where the prior does influence
the posterior distribution.
接著來看另一個事前信念確實影響事後分配的例子
107
00:06:28.620 --> 00:06:32.570
Here we have some belief that
the coin is probably fair.
這道事前分配表示硬幣是公正的
108
00:06:32.570 --> 00:06:36.680
We think that values around 0.5
are somewhat more likely but
設定統計值在0.5多少表達這樣的信念
109
00:06:36.680 --> 00:06:39.220
we don't have a very strong
conviction about this.
不過分配的變異表示對此信念不會過度堅持
110
00:06:40.730 --> 00:06:42.600
We again observe the same data.
獲得的資料也是10次之中擲出6次正面
111
00:06:42.600 --> 00:06:47.450
We flip the coin ten times
six out of ten the coin lands heads.
獲得的資料也是10次之中擲出6次正面
112
00:06:47.450 --> 00:06:50.170
So this is exactly the same
likelihood function that we had in
所以得到與前一個例子同樣的似然性函數
113
00:06:50.170 --> 00:06:53.770
the previous example but
we have a slightly different prior.
只是擁有不大相同的事前信念
114
00:06:53.770 --> 00:06:58.030
Combining our prior belief with
the observed data yields a posterior
兩者結合之後就能得到一道新的事後機率分配
115
00:06:58.030 --> 00:07:02.040
distribution that no longer falls exactly
on top of the likelihood function.
不過這道分配和似然性函數完全不疊合
116
00:07:03.700 --> 00:07:06.250
Indeed we see that the posterior
這道事後機率分配稍微偏向事前信念的beta分配
117
00:07:06.250 --> 00:07:10.720
is slightly shrinking towards
the prior belief that we held.
這道事後機率分配稍微偏向事前信念的beta分配
118
00:07:15.080 --> 00:07:18.850
Now
the posterior distribution
有了事前機率分配與事後率分配
119
00:07:18.850 --> 00:07:22.170
we can calculate something
that's known as a Bayes factor.
就能計算貝氏因子(Bayes Factor)
120
00:07:22.170 --> 00:07:26.590
The Bayes factor is the relative evidence
for one model compared to another model.
貝式因子表示手上證據支持兩種模型的相對程度
121
00:07:28.170 --> 00:07:30.239
Now compare this to likelihood ratios.
現在我們用似然性比值計算支持程度
122
00:07:30.239 --> 00:07:33.345
In a likelihood ratio we only
have one distribution and
透過這道似然性函數,比較兩個統計值的似然性
123
00:07:33.345 --> 00:07:37.749
we're comparing two different values
of theta on the likelihood function.
透過這道似然性函數,比較兩個統計值的似然性
124
00:07:37.749 --> 00:07:41.600
But here we'll use the likelihood
from the prior distribution and
現在要用這道函數計算事前分配的似然性
125
00:07:41.600 --> 00:07:45.900
compare it to the likelihood
of the posterior distribution.
與事後分配的似然性相互比較
126
00:07:45.900 --> 00:07:46.430
Let's take a look.
接著看要怎麼做
127
00:07:48.660 --> 00:07:51.450
We have data from 20 coin flips.
現在手上有擲20枚硬幣的資料
128
00:07:51.450 --> 00:07:52.960
They came up heads 10 times.
其中10次正面朝上
129
00:07:54.430 --> 00:07:56.670
Let's take a look at two different priors.
對此有兩種事前信念
130
00:07:56.670 --> 00:08:01.581
Either a Beta prior of (1
means it's completely uninformative
一種是不知這枚硬幣公正與否的均勻分配beta(1
131
00:08:01.581 --> 00:08:02.817
the uniform prior.
一種是不知這枚硬幣公正與否的均勻分配beta(1
132
00:08:02.817 --> 00:08:05.517
But we can also compare
it to a Beta of (4
另一種是認為這是枚公正硬幣的beta(4
133
00:08:05.517 --> 00:08:09.343
4) where we have some expectation
that this is a fair coin.
另一種是認為這是枚公正硬幣的beta(4
134
00:08:09.343 --> 00:08:13.910
Now
see the prior distribution in
這張圖中代表事前分配的灰色線正是均勻分佈
135
00:08:13.910 --> 00:08:17.810
the gray line that's
uniformly distributed.
這張圖中代表事前分配的灰色線正是均勻分佈
136
00:08:17.810 --> 00:08:20.480
We don't have very strong
prior beliefs in this case.
這表示沒有對特定結果存在強烈信念
137
00:08:20.480 --> 00:08:22.510
Everything is equally likely.
任何結果都可以接受
138
00:08:22.510 --> 00:08:26.010
Then we've collected some data
now we have a posterior distribution
收集資料後得到的事後分配
139
00:08:26.010 --> 00:08:28.859
which in this case falls exactly
on top of the likelihood function.
與由此獲得的似然性函數疊合
140
00:08:29.960 --> 00:08:33.770
You can see that there is a relative
difference of the likelihood of
這裡可以看到兩個分配在任意一點的似然性差異
141
00:08:33.770 --> 00:08:36.770
both of these functions
at specific points.
這裡可以看到兩個分配在任意一點的似然性差異
142
00:08:36.770 --> 00:08:39.970
Now
hypothesis that we're interested in.
現在必須要決定真正有興趣的假設是針對那一個值
143
00:08:39.970 --> 00:08:43.600
Let's say that we want to know
whether this is a fair coin or not.
比方說要知道這枚硬幣是否公正
144
00:08:43.600 --> 00:08:48.262
So in this case
the hypothesis that theta is 0.5.
要測試這個假設就要看theta為0.5的情況
145
00:08:48.262 --> 00:08:53.006
We can compare the prior distribution with
the posterior distribution to see how
比較事前與事後分配的差異可以了解收集的資料改變信念的程度
146
00:08:53.006 --> 00:08:56.156
much our belief has changed
by collecting some data.
比較事前與事後分配的差異可以了解收集的資料改變信念的程度
147
00:08:56.156 --> 00:09:01.830
So the prior at 0.5 was much lower
than the posterior is at 0.5.
這張圖明顯表示在0.5的事前分配似然性低於事後分配
148
00:09:01.830 --> 00:09:06.149
And the ratio of these two points is 3.7.
得到的似然性比值是3.7
149
00:09:09.470 --> 00:09:13.190
This is a slightly different version
where we only have a different prior.
這張圖的事前分配是beta(4
150
00:09:13.190 --> 00:09:17.910
We've observed exactly the same data but
the prior now is slightly more informed.
雖然資料也是相同的,但是這個事前分配包含有關這枚硬幣的信念
151
00:09:17.910 --> 00:09:20.400
We already expected that
this coin would be fair.
這個情況已事先預期這是枚公正硬幣
152
00:09:20.400 --> 00:09:23.760
We collect the same amount of data and
we again see that for
以同樣的資料計算兩個分配的似然性
153
00:09:23.760 --> 00:09:28.160
the theta of 0.5 the hypothesis
that this is a fair coin.
並以theta是0.5的似然性比檢測假設
154
00:09:28.160 --> 00:09:32.860
Our belief in the hypothesis that
the coin is fair has increased.
事後分配表示對於硬幣是公正的信念增加了
155
00:09:32.860 --> 00:09:35.070
But because we already
had quite a strong prior
不過因為事前先有特定的預期
156
00:09:35.070 --> 00:09:38.699
the data only increases
our belief with 1.9.
增加的比值只有1.9
157
00:09:41.430 --> 00:09:43.786
So after looking at the data
儘管資料呈現的似然性函數是相同的
158
00:09:43.786 --> 00:09:49.257
the hypothesis that the theta
is 0.5 has become either 1.91 or
theta為0.5的似然性比是1.91還是3.70
159
00:09:49.257 --> 00:09:54.460
3.70 times more likely
depending on the prior that we had.
完全取決於接受那種事前信念
160
00:09:55.620 --> 00:10:01.148
So we can see how base vectors tell you
something about the increased evidence or
由此可以了解貝氏因子如何告訴你增加的證據力有多少
161
00:10:01.148 --> 00:10:06.100
the belief that you have about
the specific hypothesis compared to your
或者根據事前與事後分配,推估證據比較支持那個假設
162
00:10:06.100 --> 00:10:07.932
prior and the posterior.
或者根據事前與事後分配,推估證據比較支持那個假設
163
00:10:07.932 --> 00:10:12.039
Now we can also look at Bayesian
statistics not from a hypothesis testing
除了由假設檢定的角度了解貝氏統計如何根據事前與事後分配測試假設
164
00:10:12.039 --> 00:10:16.233
viewpoint where you test the hypothesis
and the prior and the posterior.
除了由假設檢定的角度了解貝氏統計如何根據事前與事後分配測試假設
165
00:10:16.233 --> 00:10:20.990
But we just want to estimate which
values do we think is most likely.
也可以估計想測量的真實數值
166
00:10:20.990 --> 00:10:24.410
So this is not Bayesian hypothesis
testing
這樣的方法就叫做貝氏估計
167
00:10:25.700 --> 00:10:28.960
We also use the posterior
distribution to estimate.
估計同樣需要事後分配
168
00:10:30.370 --> 00:10:33.990
So instead of testing two different
models
不同於用事前與事後分配檢驗兩種模型
169
00:10:33.990 --> 00:10:38.090
we will use the posterior just
to estimate plausible values.
現在要用事後分配估計最可信的數值
170
00:10:39.170 --> 00:10:43.850
Which values do you believe are most
likely based on your prior belief and
要怎樣根據事前信念與觀察到的資料估計數值呢?
171
00:10:43.850 --> 00:10:44.910
the data you have observed?
要怎樣根據事前信念與觀察到的資料估計數值呢?
172
00:10:47.410 --> 00:10:53.510
In this graph we again see a prior
distribution and a posterior distribution.
這張圖也有之前看過的事前與事後機率分配
173
00:10:53.510 --> 00:10:57.320
And in this case
an uninformative prior
因為事前分配是無資訊的
174
00:10:57.320 --> 00:10:59.770
the posterior fall exactly
on top of each other.
似然性函數與事後分配再一次完美疊合
175
00:11:01.220 --> 00:11:05.580
We can calculate the mean of the posterior
distribution and we can calculate what
我們可以計算事後分配的平均值
176
00:11:05.580 --> 00:11:10.400
in Bayesian statistics
is known as 95% credible interval.
以及貝氏統計稱呼的95%可信區間(credible interval)
177
00:11:10.400 --> 00:11:16.260
A credible interval contains 95% of
the values that you find most plausible.
可信區間涵蓋的數值有95%的可能性是真實的
178
00:11:17.880 --> 00:11:20.530
So this is an expression of your belief.
也就是你相信為真的數值範圍
179
00:11:20.530 --> 00:11:24.620
Which values do you believe are most
likely given your prior and
範圍中那些數值是你最相信的
全由事前信念與手上資料決定
180
00:11:24.620 --> 00:11:25.410
the observed data?
範圍中那些數值是你最相信的
全由事前信念與手上資料決定
181
00:11:26.800 --> 00:11:30.230
Now
have used an uninformative prior
這個例子裡的事前信念是無資訊的
182
00:11:30.230 --> 00:11:34.820
you can see that the likelihood function
is basically everything that matters.
所以由手上資料構成的似然性函數
183
00:11:35.860 --> 00:11:38.810
The posterior distribution is
completely determined by the data.
決定事後分配的模樣
184
00:11:40.070 --> 00:11:45.860
And this is a situation where the 95%
credible interval in Bayesian statistics
只有這種狀況才會看到貝氏統計的可信區間
185
00:11:45.860 --> 00:11:51.900
exactly matches with the 95% confidence
interval in frequentist statistics.
與次數主義統計的信賴區間是完全一致的
186
00:11:51.900 --> 00:11:53.040
But that's not always the case.
不過除非所有研究的事前信念都是無資訊的
187
00:11:53.040 --> 00:11:56.890
This is only true when we
use an uninformative prior.
這種狀況一輩子很難碰到幾次
188
00:11:56.890 --> 00:11:59.960
Let's see how this changes when
we use a more informative prior.
接著來看看有資訊的事前信念會有什麼結果
189
00:12:02.030 --> 00:12:04.650
Now let's take a look at this
slightly different situation.
這張圖呈現另一種狀況的分析結果
190
00:12:04.650 --> 00:12:09.530
Rather extreme example where we had
a very strong prior belief that
這個狀況的事前信念是相信這枚硬幣有2/3的機會擲出正面
191
00:12:09.530 --> 00:12:15.238
the coin would come up heads about
two thirds of the time
這個狀況的事前信念是相信這枚硬幣有2/3的機會擲出正面
192
00:12:15.238 --> 00:12:18.530
You can see the grey distribution
peaking at this point
所以灰色的事前分配最高峰就在0.666
193
00:12:18.530 --> 00:12:22.050
this was our prior belief
before we collected some data.
表達收集資料前對硬幣的信念
194
00:12:22.050 --> 00:12:24.833
But then we collect data and we actually
find a completely different pattern.
但是收集的資料卻呈現與事前信念不同的模樣
195
00:12:24.833 --> 00:12:28.583
We see that the coin came up
heads about 40% of the time.
投擲硬幣很多次之後
196
00:12:28.583 --> 00:12:30.748
And we collected quite a lot of data
得到的似然性函數最高點出現在0.4
197
00:12:30.748 --> 00:12:33.910
the blue likelihood function
is completely different.
和事前分配完全不一樣
198
00:12:35.080 --> 00:12:38.640
We can also see that the black
posterior distribution is now
以黑色曲線標示的事後分配則落在兩者之間
199
00:12:38.640 --> 00:12:40.990
a little bit in between of
both of these functions.
以黑色曲線標示的事後分配則落在兩者之間
200
00:12:42.710 --> 00:12:47.495
So the data and the prior are merged
into a posterior distribution.
代表事前信念與手上資料融合的結果
201
00:12:47.495 --> 00:12:52.459
And the 95% credible interval
now contains values that no
95%可信區間與95%信賴區間並無重疊
202
00:12:52.459 --> 00:12:56.656
longer match within 95%
confidence interval.
95%可信區間與95%信賴區間並無重疊
203
00:12:56.656 --> 00:13:01.533
A very strong prior has made it so
that our believe in the most likely most
有特定預期的事前信念讓可信區間的範圍明顯縮小
204
00:13:01.533 --> 00:13:05.700
plausible values the values
that we feel most confident in.
有特定預期的事前信念讓可信區間的範圍明顯縮小
205
00:13:05.700 --> 00:13:08.749
Are no longer exactly matching
up with the data
也因為事前信念的影響
206
00:13:08.749 --> 00:13:12.090
they're slightly influenced
by the prior that we have.
讓可信區間的範圍與資料不完全重合
207
00:13:13.890 --> 00:13:16.790
This also shows a strength
of Bayesian statistics.
這個例子表現了貝氏統計的能耐
208
00:13:16.790 --> 00:13:20.160
We had a prior belief
quantified by putting a function on it
透過事前信念能預期最合理的統計值
209
00:13:20.160 --> 00:13:23.090
and then we observed some data.
再考慮收集的資料
210
00:13:23.090 --> 00:13:25.260
And you can see that our
belief can be changed.
接著計算事後信念改變的程度
211
00:13:25.260 --> 00:13:30.422
In this case
is still sort of halfway in between.
這個例子的事後分配恰好落在中間
212
00:13:30.422 --> 00:13:35.171
But if we collect enough data
we should see that the posterior belief
如果收集足夠的資料,就會看到事後分配相當接近似然性函數
213
00:13:35.171 --> 00:13:39.750
becomes very
very close to the likelihood function.
如果收集足夠的資料,就會看到事後分配相當接近似然性函數
214
00:13:39.750 --> 00:13:44.440
So this shows that if we collect enough
data
收集更多資料之後,對於硬幣是否公正的信念會逐漸改變
215
00:13:44.440 --> 00:13:47.253
we can become convinced
about something else.
逐漸相信真實數值不同於事前信念
216
00:13:47.253 --> 00:13:49.576
This is
a very useful way to do science.
這種推論方式非常適合科學
217
00:13:49.576 --> 00:13:53.349
And
between science and religion.
也體現科學與宗教信仰的差異
218
00:13:53.349 --> 00:13:57.536
In religion
any option to update your belief.
身為教徒者不可改變對教義的信念
219
00:13:57.536 --> 00:14:01.489
But
we have a very logical framework
貝氏統計卻提供一套有邏輯的架構
220
00:14:01.489 --> 00:14:06.091
where data can change our prior beliefs
and lead to new posterior beliefs.
能根據資料調整事前信念,邁向合理的事後信念
221
00:14:08.610 --> 00:14:15.020
So this 95% credible interval contains
the values that you find most plausible.
95%可信區間包括所有你能找到的最可能結果
222
00:14:15.020 --> 00:14:19.611
It's all about expressing and quantifying
your belief in specific values.
以具體數值表現與計量最為可信的統計值
223
00:14:19.611 --> 00:14:24.004
Now in this lecture we've seen how
can you use Bayesian statistics to
總結而言貝氏統計能量化你事先認定的合理數值
依理論收集資料,再根據資料更新合理數值的信念
224
00:14:24.004 --> 00:14:27.186
quantify your prior beliefs
collect data
總結而言貝氏統計能量化你事先認定的合理數值
依理論收集資料,再根據資料更新合理數值的信念
225
00:14:27.186 --> 00:14:31.445
then update your beliefs based on
the data that you've observed.
總結而言貝氏統計能量化你事先認定的合理數值
依理論收集資料,再根據資料更新合理數值的信念
226
00:14:31.445 --> 00:14:36.579
[MUSIC]
[課程結束]