1
00:00:00,000 --> 00:00:04,560
[MUSIC]

2
00:00:04,560 --> 00:00:05,744
And now we're back.

3
00:00:05,744 --> 00:00:09,685
So, we computed the or we defined
the likelihood function for logistic

4
00:00:09,685 --> 00:00:14,611
regression, we computed its derivative, we
talked about simple real ascent algorithm

5
00:00:14,611 --> 00:00:18,684
for finding the best parameters and
for those who went through that full,

6
00:00:18,684 --> 00:00:23,152
kind of slow, mathematical revision you
saw where that gradient came from, but

7
00:00:23,152 --> 00:00:25,090
that was totally optional.

8
00:00:25,090 --> 00:00:28,270
We can go back and
summarize where we are right now.

9
00:00:28,270 --> 00:00:29,685
So now learning problem.

10
00:00:29,685 --> 00:00:34,330
We have some training data, we threw
this some feature extractor that gets us

11
00:00:34,330 --> 00:00:37,470
H of X, and we saw a discretion model
that says the probability that it's

12
00:00:37,470 --> 00:00:43,050
a positive review is one over one
plus E to the minus W transpose H.

13
00:00:43,050 --> 00:00:46,770
And then we define the quality metric
which is the likelihood function and

14
00:00:46,770 --> 00:00:50,630
we use gradient ascent to
optimize it to get w hat.

15
00:00:50,630 --> 00:00:54,010
And the gradient descent algorithm is
pretty simple and pretty intuitive.

16
00:00:55,100 --> 00:01:03,240
Now, you're able to define what the
quality metric is for logistic regression.

17
00:01:03,240 --> 00:01:06,820
You can interpret the likelihood function
as being kind of the probability that you

18
00:01:06,820 --> 00:01:08,370
get training data right.

19
00:01:08,370 --> 00:01:10,020
And you're going to maximize that.

20
00:01:10,020 --> 00:01:12,890
We talked about the gradient
ascent algorithm that does it with

21
00:01:12,890 --> 00:01:15,710
really simple updates, and
we had this optional section where we

22
00:01:15,710 --> 00:01:18,100
derived a gradient ascent
algorithm from scratch.

23
00:01:19,550 --> 00:01:23,750
Next module, we're going to
take this one step further and

24
00:01:23,750 --> 00:01:28,070
explore the idea of regularization and
over-fitting logistic regression,

25
00:01:28,070 --> 00:01:29,860
which is a really important
thing in practice.

26
00:01:29,860 --> 00:01:31,681
So that's where we're going to
go in the next module.

27
00:01:31,681 --> 00:01:35,818
But now you're ready, even at this point,
supplement your own logistic

28
00:01:35,818 --> 00:01:39,436
regression algorithm from scratch,
which is super exciting.

29
00:01:39,436 --> 00:01:43,749
[MUSIC]