Would you like to inspect the original subtitles? These are the user uploaded subtitles that are being translated:
1
00:00:01,010 --> 00:00:03,240
Machine learning is creating
2
00:00:03,240 --> 00:00:05,430
tremendous economic value today.
3
00:00:05,430 --> 00:00:08,430
I think 99 percent of
the economic value
4
00:00:08,430 --> 00:00:09,870
created by machine
learning today
5
00:00:09,870 --> 00:00:11,955
is through one type
of machine learning,
6
00:00:11,955 --> 00:00:14,025
which is called
supervised learning.
7
00:00:14,025 --> 00:00:16,470
Let's take a look
at what that means.
8
00:00:16,470 --> 00:00:18,900
Supervised machine learning or
9
00:00:18,900 --> 00:00:21,075
more commonly,
supervised learning,
10
00:00:21,075 --> 00:00:24,480
refers to algorithms
that learn x to
11
00:00:24,480 --> 00:00:28,590
y or input to output mappings.
12
00:00:28,590 --> 00:00:31,770
The key characteristic of
supervised learning is
13
00:00:31,770 --> 00:00:33,225
that you give
14
00:00:33,225 --> 00:00:36,680
your learning algorithm
examples to learn from.
15
00:00:36,680 --> 00:00:40,730
That includes the right
answers, whereby right answer,
16
00:00:40,730 --> 00:00:45,235
I mean, the correct label
y for a given input x,
17
00:00:45,235 --> 00:00:48,915
and is by seeing
correct pairs of
18
00:00:48,915 --> 00:00:52,850
input x and desired
output label y that
19
00:00:52,850 --> 00:00:54,890
the learning algorithm
eventually learns to
20
00:00:54,890 --> 00:00:57,050
take just the input
alone without
21
00:00:57,050 --> 00:00:58,970
the output label and gives
22
00:00:58,970 --> 00:01:03,185
a reasonably accurate prediction
or guess of the output.
23
00:01:03,185 --> 00:01:05,620
Let's look at some examples.
24
00:01:05,620 --> 00:01:08,010
If the input x is
25
00:01:08,010 --> 00:01:11,985
an email and the output
y is this email,
26
00:01:11,985 --> 00:01:13,590
spam or not spam,
27
00:01:13,590 --> 00:01:16,785
this gives you your spam filter.
28
00:01:16,785 --> 00:01:21,120
Or if the input is
an audio clip and
29
00:01:21,120 --> 00:01:25,565
the algorithm's job is
output the text transcript,
30
00:01:25,565 --> 00:01:29,470
then this is speech recognition.
31
00:01:29,470 --> 00:01:32,770
Or if you want to
input English and have
32
00:01:32,770 --> 00:01:35,515
it output to
corresponding Spanish,
33
00:01:35,515 --> 00:01:37,495
Arabic, Hindi,
Chinese, Japanese,
34
00:01:37,495 --> 00:01:39,445
or something else translation,
35
00:01:39,445 --> 00:01:42,385
then that's machine translation.
36
00:01:42,385 --> 00:01:46,000
Or the most lucrative form
of supervised learning
37
00:01:46,000 --> 00:01:50,110
today is probably used
in online advertising.
38
00:01:50,110 --> 00:01:53,560
Nearly all the large
online ad platforms have
39
00:01:53,560 --> 00:01:56,950
a learning algorithm that
inputs some information about
40
00:01:56,950 --> 00:01:59,815
an ad and some
information about you
41
00:01:59,815 --> 00:02:01,825
and then tries to figure out
42
00:02:01,825 --> 00:02:04,120
if you will click
on that ad or not.
43
00:02:04,120 --> 00:02:05,970
Because by showing
you ads they're
44
00:02:05,970 --> 00:02:07,970
just slightly more
likely to click on,
45
00:02:07,970 --> 00:02:09,910
for these large
online ad platforms,
46
00:02:09,910 --> 00:02:11,710
every click is revenue,
47
00:02:11,710 --> 00:02:13,360
this actually drives a lot of
48
00:02:13,360 --> 00:02:15,070
revenue for these companies.
49
00:02:15,070 --> 00:02:18,055
This is something I once
done a lot of work on,
50
00:02:18,055 --> 00:02:20,665
maybe not the most
inspiring application,
51
00:02:20,665 --> 00:02:23,560
but it certainly has a
significant economic impact
52
00:02:23,560 --> 00:02:25,195
in some countries today.
53
00:02:25,195 --> 00:02:28,870
Or if you want to build
a self-driving car,
54
00:02:28,870 --> 00:02:31,180
the learning algorithm
would take as input
55
00:02:31,180 --> 00:02:33,715
an image and some
information from
56
00:02:33,715 --> 00:02:35,665
other sensors such as a radar or
57
00:02:35,665 --> 00:02:39,775
other things and then try
to output the position of,
58
00:02:39,775 --> 00:02:41,320
say, other cars so that
59
00:02:41,320 --> 00:02:42,700
your self-driving car can
60
00:02:42,700 --> 00:02:45,230
safely drive around
the other cars.
61
00:02:45,230 --> 00:02:47,565
Or take manufacturing.
62
00:02:47,565 --> 00:02:49,190
I've actually done
a lot of work in
63
00:02:49,190 --> 00:02:51,835
this sector at learning AI.
64
00:02:51,835 --> 00:02:54,050
You can have a learning
algorithm takes as
65
00:02:54,050 --> 00:02:57,395
input a picture of a
manufactured product,
66
00:02:57,395 --> 00:03:01,415
say a cell phone that just
rolled off the production line
67
00:03:01,415 --> 00:03:03,680
and have the learning
algorithm output
68
00:03:03,680 --> 00:03:05,845
whether or not
there is a scratch,
69
00:03:05,845 --> 00:03:08,715
dent, or other defect
in the product.
70
00:03:08,715 --> 00:03:11,600
This is called visual
inspection and it's helping
71
00:03:11,600 --> 00:03:13,340
manufacturers reduce or prevent
72
00:03:13,340 --> 00:03:15,410
defects in their products.
73
00:03:15,410 --> 00:03:17,750
In all of these applications,
74
00:03:17,750 --> 00:03:20,465
you will first train your
model with examples of
75
00:03:20,465 --> 00:03:23,480
inputs x and the right answers,
76
00:03:23,480 --> 00:03:25,670
that is the labels y.
77
00:03:25,670 --> 00:03:28,340
After the model has
learned from these input,
78
00:03:28,340 --> 00:03:30,365
output, or x and y pairs,
79
00:03:30,365 --> 00:03:33,065
they can then take a
brand new input x,
80
00:03:33,065 --> 00:03:34,505
something it has
never seen before,
81
00:03:34,505 --> 00:03:36,110
and try to produce the
82
00:03:36,110 --> 00:03:39,815
appropriate
corresponding output y.
83
00:03:39,815 --> 00:03:44,420
Let's dive more deeply
into one specific example.
84
00:03:44,420 --> 00:03:46,565
Say you want to predict
85
00:03:46,565 --> 00:03:49,730
housing prices based on
the size of the house.
86
00:03:49,730 --> 00:03:52,280
You've collected
some data and say
87
00:03:52,280 --> 00:03:55,390
you plot the data and
it looks like this.
88
00:03:55,390 --> 00:03:57,485
Here on the horizontal axis
89
00:03:57,485 --> 00:03:59,825
is the size of the
house in square feet.
90
00:03:59,825 --> 00:04:01,730
Yes, I live in the United States
91
00:04:01,730 --> 00:04:03,155
where we still use square feet.
92
00:04:03,155 --> 00:04:06,160
I know most of the world
uses square meters.
93
00:04:06,160 --> 00:04:10,430
Here on the vertical axis is
the price of the house in,
94
00:04:10,430 --> 00:04:12,820
say, thousands of dollars.
95
00:04:12,820 --> 00:04:16,850
With this data, let's say a
friend wants to know what's
96
00:04:16,850 --> 00:04:21,095
the price for their
750 square foot house.
97
00:04:21,095 --> 00:04:23,485
How can the learning
algorithm help you?
98
00:04:23,485 --> 00:04:25,130
One thing a learning algorithm
99
00:04:25,130 --> 00:04:26,615
might be able to do is say,
100
00:04:26,615 --> 00:04:28,280
for the straight line to
101
00:04:28,280 --> 00:04:31,700
the data and reading
off the straight line,
102
00:04:31,700 --> 00:04:33,350
it looks like your
friend's house could
103
00:04:33,350 --> 00:04:35,225
be sold for maybe about,
104
00:04:35,225 --> 00:04:38,320
I don't know, $150,000.
105
00:04:38,320 --> 00:04:40,670
But fitting a
straight line isn't
106
00:04:40,670 --> 00:04:42,830
the only learning
algorithm you can use.
107
00:04:42,830 --> 00:04:44,390
There are others that could work
108
00:04:44,390 --> 00:04:46,280
better for this application.
109
00:04:46,280 --> 00:04:49,160
For example, routed and
fitting a straight line,
110
00:04:49,160 --> 00:04:52,460
you might decide that it's
better to fit a curve,
111
00:04:52,460 --> 00:04:54,320
a function that's slightly more
112
00:04:54,320 --> 00:04:58,010
complicated or more complex
than a straight line.
113
00:04:58,010 --> 00:05:00,950
If you do that and make
a prediction here,
114
00:05:00,950 --> 00:05:02,645
then it looks like, well,
115
00:05:02,645 --> 00:05:07,725
your friend's house could be
sold for closer to $200,000.
116
00:05:07,725 --> 00:05:11,225
One of the things you
see later in this class
117
00:05:11,225 --> 00:05:14,240
is how you can decide whether
to fit a straight line,
118
00:05:14,240 --> 00:05:15,620
a curve, or
119
00:05:15,620 --> 00:05:20,310
another function that is even
more complex to the data.
120
00:05:20,310 --> 00:05:22,870
Now, it doesn't seem
appropriate to pick
121
00:05:22,870 --> 00:05:25,440
the one that gives your
friend the best price,
122
00:05:25,440 --> 00:05:27,790
but one thing you
see is how to get
123
00:05:27,790 --> 00:05:30,400
an algorithm to systematically
124
00:05:30,400 --> 00:05:32,680
choose the most
appropriate line or
125
00:05:32,680 --> 00:05:36,445
curve or other thing
to fit to this data.
126
00:05:36,445 --> 00:05:38,770
What you've seen
in this slide is
127
00:05:38,770 --> 00:05:41,335
an example of
supervised learning.
128
00:05:41,335 --> 00:05:43,660
Because we gave the
algorithm a dataset in
129
00:05:43,660 --> 00:05:45,910
which the so-called
right answer,
130
00:05:45,910 --> 00:05:47,980
that is the label or
131
00:05:47,980 --> 00:05:52,220
the correct price y is given
for every house on the plot.
132
00:05:52,220 --> 00:05:54,820
The task of the learning
algorithm is to
133
00:05:54,820 --> 00:05:57,295
produce more of
these right answers,
134
00:05:57,295 --> 00:05:59,440
specifically predicting what is
135
00:05:59,440 --> 00:06:00,640
the likely price for
136
00:06:00,640 --> 00:06:03,445
other houses like
your friend's house.
137
00:06:03,445 --> 00:06:06,275
That's why this is
supervised learning.
138
00:06:06,275 --> 00:06:08,750
To define a little
bit more terminology,
139
00:06:08,750 --> 00:06:10,940
this housing price prediction is
140
00:06:10,940 --> 00:06:12,920
the particular type of
supervised learning
141
00:06:12,920 --> 00:06:14,635
called regression.
142
00:06:14,635 --> 00:06:17,030
By regression, I
mean we're trying
143
00:06:17,030 --> 00:06:19,115
to predict a number from
144
00:06:19,115 --> 00:06:21,410
infinitely many possible numbers
145
00:06:21,410 --> 00:06:24,590
such as the house
prices in our example,
146
00:06:24,590 --> 00:06:28,035
which could be 150,000 or
147
00:06:28,035 --> 00:06:33,645
70,000 or 183,000 or any
other number in between.
148
00:06:33,645 --> 00:06:37,020
That's supervised
learning, learning input,
149
00:06:37,020 --> 00:06:39,620
output, or x to y mappings.
150
00:06:39,620 --> 00:06:42,020
You saw in this
video an example of
151
00:06:42,020 --> 00:06:45,620
regression where the task
is to predict number.
152
00:06:45,620 --> 00:06:48,845
But there's also a
second major type of
153
00:06:48,845 --> 00:06:52,190
supervised learning problem
called classification.
154
00:06:52,190 --> 00:06:55,890
Let's take a look at what
that means in the next video.11048
Can't find what you're looking for?
Get subtitles in any language from opensubtitles.com, and translate them here.