1
00:00:00,820 --> 00:00:02,970
- [Speaker] Next, let's
take a look at a demo of

2
00:00:02,970 --> 00:00:06,410
Watson's visual recognition capabilities.

3
00:00:06,410 --> 00:00:07,950
And as you're going to see,

4
00:00:07,950 --> 00:00:10,410
they provide both pretrained models

5
00:00:10,410 --> 00:00:12,450
and you can train your own models

6
00:00:12,450 --> 00:00:16,470
for recognizing things in images in video.

7
00:00:16,470 --> 00:00:19,010
So the pretrained models
are able to do things like

8
00:00:19,010 --> 00:00:22,030
colors and objects and faces, text, food,

9
00:00:22,030 --> 00:00:24,090
inappropriate content and more.

10
00:00:24,090 --> 00:00:27,220
And the training of your own models

11
00:00:27,220 --> 00:00:31,230
would be based on whatever
applications you need to develop.

12
00:00:31,230 --> 00:00:33,230
So I'll show you the
demo in just a moment.

13
00:00:33,230 --> 00:00:35,850
But the reason this is interesting is that

14
00:00:35,850 --> 00:00:39,720
one of the key areas in
which machine learning

15
00:00:39,720 --> 00:00:44,720
and deep learning are used is
computer vision applications.

16
00:00:45,520 --> 00:00:48,510
So for instance, in both of
the next couple of lessons,

17
00:00:48,510 --> 00:00:50,640
one of the examples that we'll look at

18
00:00:50,640 --> 00:00:52,250
in each of those lessons

19
00:00:52,250 --> 00:00:55,140
is going to be where we analyze

20
00:00:55,140 --> 00:00:58,040
images of handwritten digits

21
00:00:58,040 --> 00:01:00,550
to try to figure out whether a given image

22
00:01:00,550 --> 00:01:03,090
is a zero, a one, a two, etc.,

23
00:01:03,090 --> 00:01:04,700
all the way up through nine.

24
00:01:04,700 --> 00:01:06,480
This is something for instance

25
00:01:06,480 --> 00:01:09,050
that the post office needs to do

26
00:01:09,050 --> 00:01:11,060
in order to look at zip codes

27
00:01:11,060 --> 00:01:14,590
and in a computerized and mechanical way

28
00:01:14,590 --> 00:01:17,380
route mail to the appropriate locations

29
00:01:17,380 --> 00:01:19,580
for example throughout the United States.

30
00:01:19,580 --> 00:01:22,380
So here we have a prebundled service

31
00:01:22,380 --> 00:01:24,300
that knows how to do things like this

32
00:01:24,300 --> 00:01:28,130
but we're also going to
demonstrate using Python

33
00:01:28,130 --> 00:01:31,060
and machine learning and
deep learning libraries

34
00:01:31,060 --> 00:01:33,170
how you can create your own models

35
00:01:33,170 --> 00:01:35,810
that are actually able to predict things

36
00:01:35,810 --> 00:01:38,170
like what digit an image is

37
00:01:38,170 --> 00:01:42,150
with over 99% accuracy,
which is just phenomenal.

38
00:01:42,150 --> 00:01:44,330
And as you'll see, it's only going to take

39
00:01:44,330 --> 00:01:47,840
a couple of lines of code
to build those models.

40
00:01:47,840 --> 00:01:50,523
So let me switch over to
this demo for a moment here.

41
00:01:51,490 --> 00:01:54,890
And by the way, when you
first come up to this page,

42
00:01:54,890 --> 00:01:57,290
it's going to show at the
top of the page like this,

43
00:01:57,290 --> 00:01:58,910
I had already scrolled down.

44
00:01:58,910 --> 00:02:01,030
But in the demo they start out

45
00:02:01,030 --> 00:02:03,290
with a custom model here

46
00:02:03,290 --> 00:02:08,140
and this particular model
is for an insurance company.

47
00:02:08,140 --> 00:02:10,850
So they're looking at insurance claims

48
00:02:10,850 --> 00:02:12,130
and they have images of things

49
00:02:12,130 --> 00:02:14,880
and they want to be able
to analyze those images

50
00:02:14,880 --> 00:02:17,790
quickly to figure out what
type of problem it is.

51
00:02:17,790 --> 00:02:20,710
So clearly this is an image of a flat tire

52
00:02:20,710 --> 00:02:23,330
and when they ran this image through

53
00:02:23,330 --> 00:02:25,490
the custom insurance classifier,

54
00:02:25,490 --> 00:02:28,900
it classified this with 91% certainty

55
00:02:28,900 --> 00:02:30,800
as being a flat tire.

56
00:02:30,800 --> 00:02:31,970
Now down at the bottom here,

57
00:02:31,970 --> 00:02:34,130
they have other images that you can try.

58
00:02:34,130 --> 00:02:36,350
So for instance if I click this one,

59
00:02:36,350 --> 00:02:38,630
it looks like somebody vandalized this van

60
00:02:38,630 --> 00:02:41,990
so there's a 64% likelihood
based on this image

61
00:02:41,990 --> 00:02:43,770
that the issue is vandalism.

62
00:02:43,770 --> 00:02:46,530
But if you look closely
at the bottom left here,

63
00:02:46,530 --> 00:02:49,700
because of the fact that
this van is parked at a curb

64
00:02:49,700 --> 00:02:53,230
it kind of looks like
that wheel or that tire

65
00:02:53,230 --> 00:02:56,820
may be flat as well so
it's also giving a 53%

66
00:02:56,820 --> 00:03:01,590
likelihood that there is
a flat tire in this case.

67
00:03:01,590 --> 00:03:04,890
And clearly this one, once
again, is a flat tire.

68
00:03:04,890 --> 00:03:06,250
This one, if I click on it,

69
00:03:06,250 --> 00:03:08,150
looks like there was
a motorcycle accident.

70
00:03:08,150 --> 00:03:10,300
So 90% certainty on that.

71
00:03:10,300 --> 00:03:12,820
And there's additional
images that you can play with

72
00:03:12,820 --> 00:03:14,740
down here for this custom model.

73
00:03:14,740 --> 00:03:18,053
And then separately, if I go
to the pretrained models area,

74
00:03:19,470 --> 00:03:21,920
you'll notice in this
image it's giving back

75
00:03:21,920 --> 00:03:23,610
a lot of information here.

76
00:03:23,610 --> 00:03:28,110
So with 96% certainty,
there's fabric in this image.

77
00:03:28,110 --> 00:03:30,500
And in this case it's recognizing

78
00:03:30,500 --> 00:03:34,630
that this is a Harris Tweed
jacket with 87% likelihood

79
00:03:34,630 --> 00:03:37,530
and that there's clothing
in here with 80% likelihood.

80
00:03:37,530 --> 00:03:41,230
And lots of other possibilities as well.

81
00:03:41,230 --> 00:03:44,960
There's other, this is the
general model by the way.

82
00:03:44,960 --> 00:03:48,080
Other pretrained models that
they have are for faces,

83
00:03:48,080 --> 00:03:52,530
food, explicit content and text as well.

84
00:03:52,530 --> 00:03:55,340
So depending on what
image you're looking at,

85
00:03:55,340 --> 00:03:58,170
you may want to look at the
results from different models.

86
00:03:58,170 --> 00:04:00,930
So for this, clearly there's a face here,

87
00:04:00,930 --> 00:04:03,390
it's recognizing that
there's a person in here.

88
00:04:03,390 --> 00:04:05,960
It's saying with 75% likelihood.

89
00:04:05,960 --> 00:04:07,833
If I go to the face model,

90
00:04:09,230 --> 00:04:11,970
it's saying not only that there's a face

91
00:04:11,970 --> 00:04:15,000
but specifically that this person

92
00:04:15,000 --> 00:04:18,340
is probably between the ages of 19 and 22

93
00:04:18,340 --> 00:04:20,520
with 93% certainty.

94
00:04:20,520 --> 00:04:23,470
And they're saying with 100% certainty

95
00:04:23,470 --> 00:04:25,200
that this is a female.

96
00:04:25,200 --> 00:04:28,250
And of course if you use a
model that's not intended

97
00:04:28,250 --> 00:04:31,890
you're not necessarily going
to get the results you expect.

98
00:04:31,890 --> 00:04:34,730
It is clearly recognizing
that this is not food

99
00:04:34,730 --> 00:04:38,470
with 100% certainty in this case.

100
00:04:38,470 --> 00:04:42,220
So again, pretrained models
for lots of existing cases.

101
00:04:42,220 --> 00:04:44,670
Which may, in fact, be all you need

102
00:04:44,670 --> 00:04:46,210
for the apps that you're building.

103
00:04:46,210 --> 00:04:48,130
But you also have the ability to build

104
00:04:48,130 --> 00:04:49,570
your own custom models.

105
00:04:49,570 --> 00:04:52,520
So if I'm working for an insurance company

106
00:04:52,520 --> 00:04:56,150
and I have tens of millions of photos

107
00:04:56,150 --> 00:04:59,380
of insurance claim issues over the years,

108
00:04:59,380 --> 00:05:02,540
I can train models based on those images

109
00:05:02,540 --> 00:05:04,840
and then hopefully make it easier

110
00:05:04,840 --> 00:05:08,520
for my insurance adjusters to categorize

111
00:05:08,520 --> 00:05:10,563
future claims as they come in.