1
00:00:00,000 --> 00:00:04,733
[MUSIC]

2
00:00:04,733 --> 00:00:08,339
Now that we've gone through even more
material on clustering, let's spend

3
00:00:08,339 --> 00:00:11,900
a little bit of time describing what
we didn't cover in this course.

4
00:00:11,900 --> 00:00:12,750
So in this course,

5
00:00:12,750 --> 00:00:15,970
we tried to focus on the methods that
we believe are the most practical and

6
00:00:15,970 --> 00:00:21,410
most widely used tools out there for
performing retrieval and clustering.

7
00:00:21,410 --> 00:00:24,710
But there were still some other
important topics that we didn't cover in

8
00:00:24,710 --> 00:00:25,870
this course.

9
00:00:25,870 --> 00:00:27,240
So for example, in retrieval,

10
00:00:27,240 --> 00:00:30,500
there are lots of other distance
metrics that we could have described.

11
00:00:30,500 --> 00:00:32,400
We listed some of these in a module, but

12
00:00:32,400 --> 00:00:35,980
we didn't actually describe
these in any detail.

13
00:00:35,980 --> 00:00:39,620
Another thing we didn't cover was
something called distance metrics learning

14
00:00:39,620 --> 00:00:43,990
where these are procedures that
allow you to actually learn

15
00:00:43,990 --> 00:00:46,790
metrics that are useful for
the task at hand.

16
00:00:46,790 --> 00:00:51,050
Then for clustering, we didn't talk about
things like nonparametric clustering.

17
00:00:51,050 --> 00:00:53,960
Actually our hierarchical
clustering methods can be used for

18
00:00:53,960 --> 00:00:57,070
nonparametric clustering, but
we didn't describe this very explicitly.

19
00:00:57,070 --> 00:01:03,360
So nonparametric clustering, these are
methods where the complexity of the models

20
00:01:03,360 --> 00:01:08,440
or the description of the clustering can
grow as you get more and more data points.

21
00:01:08,440 --> 00:01:13,120
Another method is called spectral
clustering that can be robust to different

22
00:01:13,120 --> 00:01:18,790
cluster shapes like the Swiss roll type
images we showed in the second module,

23
00:01:18,790 --> 00:01:20,210
when we're motivating clustering and

24
00:01:20,210 --> 00:01:23,700
talking about some of the limitations of
the methods we were going to describe,

25
00:01:23,700 --> 00:01:27,440
where we have to actually specify
the shape of the clusters.

26
00:01:27,440 --> 00:01:30,010
But unfortunately spectral
clustering methods don't

27
00:01:30,010 --> 00:01:34,280
tend to have very good scalability
properties to large data sets.

28
00:01:34,280 --> 00:01:38,580
And then there are a set of ideas that are
related to things that we talked about but

29
00:01:38,580 --> 00:01:42,280
not specifically for
retrieval or clustering.

30
00:01:42,280 --> 00:01:46,590
So, as an example, mixtures of
Gaussians are commonly used for

31
00:01:46,590 --> 00:01:48,348
something called density estimation.

32
00:01:48,348 --> 00:01:52,720
So we showed this picture of this
histogram over the blue intensity of

33
00:01:52,720 --> 00:01:54,990
all the images in our data set and

34
00:01:54,990 --> 00:01:59,910
we talked about the distribution
on those intensities.

35
00:01:59,910 --> 00:02:04,040
So we can think of that as a density
over intensities, and try and

36
00:02:04,040 --> 00:02:07,950
estimate the form of that
density explicitly, rather than

37
00:02:07,950 --> 00:02:12,440
thinking about mixtures of Gaussians
as a means of clustering data points.

38
00:02:12,440 --> 00:02:15,519
So they're hand in hand,
the same tool, but different tasks.

39
00:02:16,570 --> 00:02:18,940
And then within the context
of density estimation,

40
00:02:18,940 --> 00:02:22,270
you can start talking about
things like anomaly detection,

41
00:02:22,270 --> 00:02:27,320
where this is a question of you
get some new data point, and

42
00:02:27,320 --> 00:02:31,410
does it look significantly different than
the data points that you've seen so far.

43
00:02:31,410 --> 00:02:35,355
But with what you've learned in this
class you have the tools to go out and

44
00:02:35,355 --> 00:02:36,917
learn these other methods.

45
00:02:36,917 --> 00:02:41,219
[MUSIC]