0
1
00:00:00,580 --> 00:00:06,140
In the last lesson, we spoke about supervised learning as a form of machine learning.
1

2
00:00:06,190 --> 00:00:13,270
So this is where when you're training up your computational model, you're providing labeled data. So your
2

3
00:00:13,270 --> 00:00:20,290
data is clean and everything that you feed into the model has a label that the computer can understand.
3

4
00:00:20,290 --> 00:00:26,710
So, for example, you could feed in loads and loads of pictures of flowers and each flower has a string
4

5
00:00:26,980 --> 00:00:28,950
that labels what it is,
5

6
00:00:28,960 --> 00:00:32,430
so roses or chrysanthemums, or whichever.
6

7
00:00:32,440 --> 00:00:38,440
Now, we're going to talk about unsupervised learning and this is when the teacher goes out of the room
7

8
00:00:38,530 --> 00:00:40,760
and all the children go crazy.
8

9
00:00:40,810 --> 00:00:41,710
No, I'm just kidding.
9

10
00:00:41,710 --> 00:00:49,660
It's one of the other major types of machine learning. Now, with unsupervised learning, instead of providing
10

11
00:00:49,660 --> 00:00:55,440
training data where each and every piece of data that you feed into the model is clearly labeled,
11

12
00:00:55,450 --> 00:01:01,870
instead, you're just feeding it a whole bunch of data, and you want the model to work out some sense in
12

13
00:01:01,870 --> 00:01:08,800
that data set. So it learns to find structure in your unstructured data.
13

14
00:01:08,800 --> 00:01:14,950
So in unsupervised learning, you might have a whole bunch of data that looks like this. And you're not
14

15
00:01:14,950 --> 00:01:22,000
really telling the computer what all the data means, but instead, you want the computer to simply learn
15

16
00:01:22,210 --> 00:01:29,800
and be able to see this data seems to be organized in a particular fashion, and you can actually divide
16

17
00:01:29,800 --> 00:01:32,220
them roughly along this line,
17

18
00:01:32,500 --> 00:01:39,160
and the computer can see the structure within the data so being able to differentiate that this might
18

19
00:01:39,160 --> 00:01:44,740
be two separate groups of data, having something more in common with each other than the other group.
19

20
00:01:44,740 --> 00:01:51,610
So one type of unsupervised learning is called clustering and this is where the computer learns and
20

21
00:01:51,610 --> 00:01:58,900
looks at the data and all of its various features, and figures out that all the data belongs in separate
21

22
00:01:59,170 --> 00:02:04,480
clusters, instead of a across-the-board random set of crosses.
22

23
00:02:04,480 --> 00:02:10,240
We actually have discrete clusters that probably can be categorized as grouped in clusters.
23

24
00:02:10,270 --> 00:02:12,860
So let me illustrate this with a real-life example.
24

25
00:02:12,880 --> 00:02:14,620
I think you'll make it a lot more clear.
25

26
00:02:14,980 --> 00:02:19,420
Now, let's say that you work on the engineering team at Facebook.
26

27
00:02:19,540 --> 00:02:26,380
Now, one of the problems that you're tasked with solving is, how do you get more users to engage more
27

28
00:02:26,380 --> 00:02:27,520
on Facebook?
28

29
00:02:27,520 --> 00:02:32,020
This is the dilemma that real-life Facebook engineers are faced with every day.
29

30
00:02:32,230 --> 00:02:38,050
So one of the ways of increasing engagement is by putting people in touch with friends that they didn't
30

31
00:02:38,050 --> 00:02:42,380
know were on Facebook or friends that they have an added as a friend yet.
31

32
00:02:42,670 --> 00:02:46,500
So how can you do that if they're not searching for those friends?
32

33
00:02:46,510 --> 00:02:54,520
Well, if you look into the data on Facebook, you might be able to see that this person is connected to
33

34
00:02:54,520 --> 00:02:55,390
this person.
34

35
00:02:55,690 --> 00:03:01,360
So that could be through the fact that they have both added each other as friend or maybe, you know,
35

36
00:03:01,390 --> 00:03:09,920
the guy on the left has tagged the woman on the right or various other activities that show this connection.
36

37
00:03:09,940 --> 00:03:16,630
Now, if you map this out to all the users, you'll be able to see all of their interconnected relationships
37

38
00:03:17,080 --> 00:03:20,580
and you can see which users are linked to whom.
38

39
00:03:20,590 --> 00:03:27,940
Now, even as we look at this using the human eye, you can sort of see clusters appearing, right?
39

40
00:03:27,940 --> 00:03:31,840
So maybe this is a cluster or a friendship group
40

41
00:03:32,260 --> 00:03:34,960
and this is another cluster friendship group.
41

42
00:03:34,960 --> 00:03:42,160
So using clustering and using unsupervised learning, just by looking at the Facebook raw data, the machine
42

43
00:03:42,160 --> 00:03:49,030
learning algorithm is able to tell who probably belongs in certain social groups all without any sort
43

44
00:03:49,030 --> 00:03:50,070
of training data.
44

45
00:03:50,080 --> 00:03:56,800
You know I didn't say that these and these friends are in my close inner circle or those and those friends
45

46
00:03:56,830 --> 00:03:59,920
are in my basketball team or whichever,
46

47
00:03:59,920 --> 00:04:06,690
but all of this data is worked out based on these patterns that are recognized by the computer.
47

48
00:04:06,940 --> 00:04:13,360
So what can you do using clustering? What is a application of this type of machine learning?
48

49
00:04:13,360 --> 00:04:17,050
Well, let's zoom in on this particular social circle.
49

50
00:04:17,140 --> 00:04:23,720
Now, you can see that this lady here is not yet linked to these two people, right?
50

51
00:04:23,920 --> 00:04:31,750
Even though she seems to belong in the same social group. So she probably knows those two people, but
51

52
00:04:31,750 --> 00:04:34,380
just hasn't yet added them as a friend.
52

53
00:04:34,390 --> 00:04:39,970
So if you go onto Facebook, one of the most prominent things that you see nowadays at the top is this
53

54
00:04:39,970 --> 00:04:47,800
bar where it says, "People You May Know," and the Facebook machine learning algorithm is using clustering
54

55
00:04:47,830 --> 00:04:55,420
and using unsupervised learning to look at these maps of social networks figuring out who belongs in
55

56
00:04:55,420 --> 00:05:02,330
which social circles and the people that you might know who are also in the same circle as you, but you
56

57
00:05:02,330 --> 00:05:04,920
haven't yet added them as a friend.
57

58
00:05:04,970 --> 00:05:10,140
So this is how it figures out some of those people who you might want to add as friends.
58

59
00:05:10,220 --> 00:05:13,940
Now, if you add more people as friends, then you have more interactions,
59

60
00:05:14,030 --> 00:05:17,850
you see more of their pictures, increasing engagement on Facebook.
60

61
00:05:17,900 --> 00:05:18,670
And there we go,
61

62
00:05:18,680 --> 00:05:21,330
job done using machine learning.
62

63
00:05:21,350 --> 00:05:28,040
So this is called clustering and it is a type of unsupervised machine learning.
63

64
00:05:28,070 --> 00:05:32,990
Now, there's quite a few other types of unsupervised machine learning, but it's beyond the scope of this
64

65
00:05:32,990 --> 00:05:39,920
talk at the moment. We're currently working hard on entire suite of machine learning lessons in order
65

66
00:05:39,920 --> 00:05:46,220
to teach machine learning in a way that's easily understood and we can go into deeper topics like reinforcement
66

67
00:05:46,220 --> 00:05:49,850
learning, back propagation, and lots, lots more.
67

68
00:05:49,850 --> 00:05:56,990
But for now, we've spoken about the two major types of machine learning which are supervised learning
68

69
00:05:57,890 --> 00:06:00,680
and unsupervised learning.
69

70
00:06:00,680 --> 00:06:05,240
Now, in the next lesson, I'm going to talk to you about one of my favorite types of deep learning which
70

71
00:06:05,300 --> 00:06:07,270
is reinforcement learning.
71

72
00:06:07,700 --> 00:06:09,320
So I'll see you on the next lesson.