1
00:00:00,000 --> 00:00:05,051
[MUSIC]

2
00:00:05,051 --> 00:00:09,451
So congratulations, you guys have made
it to the end of the last module of this

3
00:00:09,451 --> 00:00:13,647
course and so we've covered lots and
lots of material in this course and in

4
00:00:13,647 --> 00:00:18,410
particular the last couple of modules have
been quite advanced and quite intense.

5
00:00:19,610 --> 00:00:24,480
So in this module, in particular,
we covered Latent Dirichlet Allocation, or

6
00:00:24,480 --> 00:00:27,590
LDA, and Gibbs sampling.

7
00:00:27,590 --> 00:00:31,608
So LDA is a really widely used tool for
mixed membership modeling and

8
00:00:31,608 --> 00:00:35,203
text corpora, but variance of
the LDA model could be used for

9
00:00:35,203 --> 00:00:39,523
mixed membership modeling and
a huge range of different applications.

10
00:00:39,523 --> 00:00:43,750
And we also covered Gibbs sampling, which
is the most widely used algorithm for

11
00:00:43,750 --> 00:00:45,920
Bayesian inference.

12
00:00:45,920 --> 00:00:49,090
We specifically examined it
in the context of LDA, but

13
00:00:49,090 --> 00:00:52,540
it tends to be the most straightforward
algorithm to think about

14
00:00:52,540 --> 00:00:55,540
deriving the updates for
in any Bayesian model.

15
00:00:56,820 --> 00:00:59,660
That said, it's not always
the most scalable algorithm,

16
00:00:59,660 --> 00:01:04,810
especially just the first implementation
you might think of writing down,

17
00:01:04,810 --> 00:01:07,290
but there's a lot of work in a community.

18
00:01:07,290 --> 00:01:10,110
I'm thinking about scaling
up Gibbs sampling or

19
00:01:10,110 --> 00:01:14,375
Gibbs sampling type algorithms to really
large data sets and really big models.

20
00:01:15,655 --> 00:01:19,325
So in conclusion, I want to leave you with
a list of things that you should be able

21
00:01:19,325 --> 00:01:22,365
to do now that you've
completed this module.

22
00:01:22,365 --> 00:01:25,535
So take a little bit of time
to reflect on this list and

23
00:01:25,535 --> 00:01:28,760
think about all the advances
that you've made.

24
00:01:28,760 --> 00:01:32,390
And finally, I want to give
a big thanks to David Mimno for

25
00:01:32,390 --> 00:01:37,103
providing an outline of the example
that we walked through when we talked

26
00:01:37,103 --> 00:01:39,750
about collapse Gibbs sampling and LDA.

27
00:01:39,750 --> 00:01:43,999
[MUSIC]