Would you like to inspect the original subtitles? These are the user uploaded subtitles that are being translated:
1
00:00:03,220 --> 00:00:08,440
the rise of AI has been largely driven
2
00:00:05,560 --> 00:00:11,170
by one too in AI called machine learning
3
00:00:08,440 --> 00:00:12,849
in this video you learn what is machine
4
00:00:11,170 --> 00:00:14,500
learning so that by the end you hope
5
00:00:12,849 --> 00:00:16,299
you'll be able to start thinking how
6
00:00:14,500 --> 00:00:18,940
machine learning might be applied to
7
00:00:16,299 --> 00:00:20,859
your company or to your industry the
8
00:00:18,940 --> 00:00:24,810
most commonly used type of machine
9
00:00:20,859 --> 00:00:31,240
learning as a type of AI that learns a
10
00:00:24,810 --> 00:00:34,000
to be or input to output mappings and
11
00:00:31,240 --> 00:00:37,210
this is called supervised learning let's
12
00:00:34,000 --> 00:00:40,120
see some examples if the input a is an
13
00:00:37,210 --> 00:00:43,120
email and the output B you want is this
14
00:00:40,120 --> 00:00:46,059
email spam one on 0 1 then this is the
15
00:00:43,120 --> 00:00:49,270
core piece of AI used to build a spam
16
00:00:46,059 --> 00:00:52,420
filter or if the input is an audio clip
17
00:00:49,270 --> 00:00:54,010
and the a eyes job is output D text
18
00:00:52,420 --> 00:00:57,760
transcript dentists is speech
19
00:00:54,010 --> 00:01:00,100
recognition more examples if you want to
20
00:00:57,760 --> 00:01:01,840
input English and have it outputs a
21
00:01:00,100 --> 00:01:04,360
different language Chinese Spanish
22
00:01:01,840 --> 00:01:07,360
something else then this is machine
23
00:01:04,360 --> 00:01:09,730
translation or the most lucrative form
24
00:01:07,360 --> 00:01:10,900
of supervised learning of this type of
25
00:01:09,730 --> 00:01:13,510
machine learning maybe online
26
00:01:10,900 --> 00:01:15,880
advertising where all the large online
27
00:01:13,510 --> 00:01:18,160
ad platforms have a piece of AI that
28
00:01:15,880 --> 00:01:21,070
inputs some information above an ad and
29
00:01:18,160 --> 00:01:23,170
some information about you and tries to
30
00:01:21,070 --> 00:01:26,080
figure out will you click on this ad or
31
00:01:23,170 --> 00:01:28,060
not and by showing you the answer you
32
00:01:26,080 --> 00:01:29,980
most likely click on this turns out to
33
00:01:28,060 --> 00:01:31,780
be very lucrative maybe not the most
34
00:01:29,980 --> 00:01:33,930
inspiring application but certainly
35
00:01:31,780 --> 00:01:36,400
having a huge economic impact today or
36
00:01:33,930 --> 00:01:38,650
if you want to build a self-driving car
37
00:01:36,400 --> 00:01:40,600
one of the key pieces of AI is in the
38
00:01:38,650 --> 00:01:42,970
IDE that teaches input an image and some
39
00:01:40,600 --> 00:01:45,760
information from the radar or from other
40
00:01:42,970 --> 00:01:47,770
sensors and outputs the position of
41
00:01:45,760 --> 00:01:49,780
other costs so your self-driving car can
42
00:01:47,770 --> 00:01:51,610
avoid the other cause or in
43
00:01:49,780 --> 00:01:53,710
manufacturing I've actually done a lot
44
00:01:51,610 --> 00:01:56,200
of work in manufacturing where you take
45
00:01:53,710 --> 00:01:58,420
as input a picture of something you've
46
00:01:56,200 --> 00:02:00,310
just manufacture such as a picture of a
47
00:01:58,420 --> 00:02:02,290
cell phone coming off an assembly line
48
00:02:00,310 --> 00:02:04,840
this is a picture of a phone another
49
00:02:02,290 --> 00:02:07,030
picture taken by a phone and you want to
50
00:02:04,840 --> 00:02:08,979
output is there a scratch was there a
51
00:02:07,030 --> 00:02:11,110
dancer as some other defects on this
52
00:02:08,979 --> 00:02:13,180
thing you've just manufactured and this
53
00:02:11,110 --> 00:02:14,170
is visual inspection which is helping
54
00:02:13,180 --> 00:02:16,840
manufacturers
55
00:02:14,170 --> 00:02:19,000
reduce or prevent defects in the things
56
00:02:16,840 --> 00:02:22,000
that they're making this type of AI
57
00:02:19,000 --> 00:02:25,000
called supervised learning just learns
58
00:02:22,000 --> 00:02:28,390
input to output or a to be mappings and
59
00:02:25,000 --> 00:02:30,550
on one hand input output ABB seems quite
60
00:02:28,390 --> 00:02:32,590
limiting but when you find the right
61
00:02:30,550 --> 00:02:35,590
application scenario this can be
62
00:02:32,590 --> 00:02:37,300
incredibly valuable now the idea of
63
00:02:35,590 --> 00:02:39,310
supervised learning has been around for
64
00:02:37,300 --> 00:02:41,800
many decades but that's really taken off
65
00:02:39,310 --> 00:02:43,239
in the last few years why is this where
66
00:02:41,800 --> 00:02:44,980
my friends asked me hey Andrew why is
67
00:02:43,239 --> 00:02:47,110
supervised learning is taking off now
68
00:02:44,980 --> 00:02:48,400
there's a picture I draw for them and I
69
00:02:47,110 --> 00:02:50,620
want to show you this picture now and
70
00:02:48,400 --> 00:02:52,600
you may be able to draw this picture for
71
00:02:50,620 --> 00:02:55,150
others that ask you the same question as
72
00:02:52,600 --> 00:02:57,489
well let's say on the horizontal axis
73
00:02:55,150 --> 00:02:59,799
you plot the amount of data you have
74
00:02:57,489 --> 00:03:02,230
Berta's so for speech recognition this
75
00:02:59,799 --> 00:03:04,480
might be the amount of audio data and
76
00:03:02,230 --> 00:03:06,580
transcripts you have in a lot of
77
00:03:04,480 --> 00:03:08,440
industries the amount of data you have
78
00:03:06,580 --> 00:03:10,989
access to has really grown over the last
79
00:03:08,440 --> 00:03:13,209
couple decades thanks to the rise in the
80
00:03:10,989 --> 00:03:15,700
Internet the rise of computers a lot of
81
00:03:13,209 --> 00:03:18,130
what used to be say pieces of paper are
82
00:03:15,700 --> 00:03:20,650
now instead recorded on a digital
83
00:03:18,130 --> 00:03:22,930
computer so we've just been getting more
84
00:03:20,650 --> 00:03:24,519
and more and more data now let's say on
85
00:03:22,930 --> 00:03:27,640
the vertical axis you plot the
86
00:03:24,519 --> 00:03:29,769
performance of an AI system it turns out
87
00:03:27,640 --> 00:03:32,470
that if you use a traditional AI system
88
00:03:29,769 --> 00:03:34,870
then the performance would grow like
89
00:03:32,470 --> 00:03:37,420
this then as you feed it more data as
90
00:03:34,870 --> 00:03:39,940
performance gets a bit better but beyond
91
00:03:37,420 --> 00:03:41,350
a certain point it did not get that much
92
00:03:39,940 --> 00:03:43,720
better so that if your speech
93
00:03:41,350 --> 00:03:45,250
recognition system did not get that much
94
00:03:43,720 --> 00:03:46,690
more accurate or your online advertising
95
00:03:45,250 --> 00:03:48,010
system didn't get that much more
96
00:03:46,690 --> 00:03:50,829
accurate than showing the most relevant
97
00:03:48,010 --> 00:03:53,109
ads even as you showed it more data AI
98
00:03:50,829 --> 00:03:55,390
has really taken off recently due to the
99
00:03:53,109 --> 00:03:57,160
rise of neuro networks and deep learning
100
00:03:55,390 --> 00:03:58,660
how to find these terms more precisely
101
00:03:57,160 --> 00:04:00,340
in later videos so don't worry too much
102
00:03:58,660 --> 00:04:02,230
about what it means but now but with
103
00:04:00,340 --> 00:04:04,359
modern AI with neural networks and deep
104
00:04:02,230 --> 00:04:07,569
learning what we saw was that if you
105
00:04:04,359 --> 00:04:09,959
train a small neural network then to
106
00:04:07,569 --> 00:04:09,959
perform
107
00:04:11,030 --> 00:04:16,130
you may have heard that data is really
108
00:04:13,850 --> 00:04:18,829
important for building AI systems but
109
00:04:16,130 --> 00:04:22,130
what is data really let's take a look
110
00:04:18,829 --> 00:04:25,790
let's look at an example of a table of
111
00:04:22,130 --> 00:04:27,560
data which we also call a data set if
112
00:04:25,790 --> 00:04:29,570
you're trying to figure out how to price
113
00:04:27,560 --> 00:04:32,750
houses that you trying to buy or sell
114
00:04:29,570 --> 00:04:35,450
you might collect a data set like this
115
00:04:32,750 --> 00:04:38,389
and this can be just a spreadsheet like
116
00:04:35,450 --> 00:04:41,030
an Excel spreadsheet of data where one
117
00:04:38,389 --> 00:04:43,490
column is the size of the house say in
118
00:04:41,030 --> 00:04:45,230
square feet or square meters and the
119
00:04:43,490 --> 00:04:47,990
second column is the price of the house
120
00:04:45,230 --> 00:04:49,910
and so if you're trying to build an AI
121
00:04:47,990 --> 00:04:52,610
system a machine learning system to help
122
00:04:49,910 --> 00:04:54,710
you set prices for houses or figure out
123
00:04:52,610 --> 00:04:56,270
of a houses price appropriately you
124
00:04:54,710 --> 00:04:59,419
might decide that the size of the house
125
00:04:56,270 --> 00:05:02,570
is a and the price of the houses B and
126
00:04:59,419 --> 00:05:06,770
have an AI system learn this input to
127
00:05:02,570 --> 00:05:08,360
output or a to be mapping now rather
128
00:05:06,770 --> 00:05:10,460
than just pricing a house based on the
129
00:05:08,360 --> 00:05:12,560
size you might say well let's also
130
00:05:10,460 --> 00:05:16,280
collect data on the number of bedrooms
131
00:05:12,560 --> 00:05:20,950
of this house in that case a can be both
132
00:05:16,280 --> 00:05:24,680
of these first two columns and B can be
133
00:05:20,950 --> 00:05:26,960
just the price of the house so given a
134
00:05:24,680 --> 00:05:28,880
table of data given the data set there's
135
00:05:26,960 --> 00:05:32,090
actually up to you up to your business
136
00:05:28,880 --> 00:05:35,780
use case to decide what is a and what is
137
00:05:32,090 --> 00:05:39,590
B data is often unique to your business
138
00:05:35,780 --> 00:05:42,050
and this is an example of a data set
139
00:05:39,590 --> 00:05:43,850
that a real estate agency might have if
140
00:05:42,050 --> 00:05:46,340
they trying to help price
141
00:05:43,850 --> 00:05:49,100
pulses and it's up to you to decide what
142
00:05:46,340 --> 00:05:51,530
is a and what is B and how to choose
143
00:05:49,100 --> 00:05:54,229
these definitions of a and B to make it
144
00:05:51,530 --> 00:05:56,990
valuable for your business as another
145
00:05:54,229 --> 00:05:59,630
example if you have a certain budget and
146
00:05:56,990 --> 00:06:01,669
you want to decide what is the size of
147
00:05:59,630 --> 00:06:04,910
house you can afford then you might
148
00:06:01,669 --> 00:06:09,229
decide that the input a is how much does
149
00:06:04,910 --> 00:06:11,560
someone spend and B is just the size of
150
00:06:09,229 --> 00:06:11,560
the home
151
00:06:12,680 --> 00:06:17,960
you might have heard terminology from AI
152
00:06:15,350 --> 00:06:20,419
such as machine learning or data science
153
00:06:17,960 --> 00:06:22,580
or neural networks or deep learning what
154
00:06:20,419 --> 00:06:24,979
do these terms mean in this video you
155
00:06:22,580 --> 00:06:27,169
see what is this terminology of the most
156
00:06:24,979 --> 00:06:28,850
important concepts of AI so that you
157
00:06:27,169 --> 00:06:30,650
will speak with others about it and
158
00:06:28,850 --> 00:06:32,840
start thinking how these things could
159
00:06:30,650 --> 00:06:35,900
apply in your business let's get started
160
00:06:32,840 --> 00:06:38,150
let's say you have a housing data set
161
00:06:35,900 --> 00:06:39,770
like this with the size of house number
162
00:06:38,150 --> 00:06:42,199
bedrooms and Rabab rooms what are the
163
00:06:39,770 --> 00:06:44,990
houses newly renovated as well as the
164
00:06:42,199 --> 00:06:48,560
price if you want to build a mobile app
165
00:06:44,990 --> 00:06:51,979
to help people price houses so this
166
00:06:48,560 --> 00:06:54,500
would be the input a and this would be
167
00:06:51,979 --> 00:06:56,660
the outputs B then this would be a
168
00:06:54,500 --> 00:06:58,460
machine learning system in particular
169
00:06:56,660 --> 00:07:01,070
it'd be one of those machine learning
170
00:06:58,460 --> 00:07:04,430
systems that learns inputs to outputs or
171
00:07:01,070 --> 00:07:08,240
a to be mappings so machine learning
172
00:07:04,430 --> 00:07:10,039
often results in a running AI system so
173
00:07:08,240 --> 00:07:11,990
there's a piece of software that any
174
00:07:10,039 --> 00:07:14,870
time of day any time of night you can
175
00:07:11,990 --> 00:07:17,419
automatically input a these properties
176
00:07:14,870 --> 00:07:20,870
of a house and a plus B so if you have
177
00:07:17,419 --> 00:07:22,159
an AI system running serving dozens or
178
00:07:20,870 --> 00:07:24,620
hundreds of thousands of millions of
179
00:07:22,159 --> 00:07:28,070
users that's usually a machine learning
180
00:07:24,620 --> 00:07:30,560
system in contrast here's something else
181
00:07:28,070 --> 00:07:33,889
you might want to do which is to have a
182
00:07:30,560 --> 00:07:36,860
team analyze your data set in order to
183
00:07:33,889 --> 00:07:38,930
gain insights so a team might come up
184
00:07:36,860 --> 00:07:41,090
with a conclusion like hey did you know
185
00:07:38,930 --> 00:07:43,759
if you have two houses of a similar size
186
00:07:41,090 --> 00:07:45,800
of a similar square footage if the house
187
00:07:43,759 --> 00:07:48,349
has three bedrooms then they cost a lot
188
00:07:45,800 --> 00:07:51,740
more than the house of two bedrooms even
189
00:07:48,349 --> 00:07:53,780
if the square footage is the same or did
190
00:07:51,740 --> 00:07:55,759
you know that newly renovated homes have
191
00:07:53,780 --> 00:07:58,520
a fifteen percent premium and this could
192
00:07:55,759 --> 00:08:00,560
help you make decisions such as given a
193
00:07:58,520 --> 00:08:02,240
similar square footage do you want to
194
00:08:00,560 --> 00:08:05,090
build a two bedroom or a three bedroom
195
00:08:02,240 --> 00:08:07,039
size in order to maximize value or is it
196
00:08:05,090 --> 00:08:08,509
worth in investments to renovate a home
197
00:08:07,039 --> 00:08:11,120
in the hope that the renovation
198
00:08:08,509 --> 00:08:14,000
increases the price you can sell a house
199
00:08:11,120 --> 00:08:17,210
for so these would be examples of data
200
00:08:14,000 --> 00:08:19,610
science projects where the output of a
201
00:08:17,210 --> 00:08:22,130
data science project is a set of
202
00:08:19,610 --> 00:08:24,860
insights that can help you make business
203
00:08:22,130 --> 00:08:26,300
decisions such as what type of house to
204
00:08:24,860 --> 00:08:28,970
build or whether to invest
205
00:08:26,300 --> 00:08:31,190
in renovation the boundaries between
206
00:08:28,970 --> 00:08:33,529
these two terms machine learning and
207
00:08:31,190 --> 00:08:35,149
data science are a little bit fuzzy and
208
00:08:33,529 --> 00:08:37,700
these terms are not used consistently
209
00:08:35,149 --> 00:08:39,769
even in industry today but what I'm
210
00:08:37,700 --> 00:08:41,839
giving here is maybe the most commonly
211
00:08:39,769 --> 00:08:44,750
used definitions of these terms but you
212
00:08:41,839 --> 00:08:48,110
will not find universal adherence to
213
00:08:44,750 --> 00:08:50,660
these definitions so formalize these two
214
00:08:48,110 --> 00:08:53,690
notions a bit more machine learning is
215
00:08:50,660 --> 00:08:54,890
the field of study that gives computers
216
00:08:53,690 --> 00:08:56,839
the ability to learn without being
217
00:08:54,890 --> 00:08:59,600
explicitly programmed this is a
218
00:08:56,839 --> 00:09:02,660
definition by author Samuel many decades
219
00:08:59,600 --> 00:09:04,370
ago after Samuel was one of the pioneers
220
00:09:02,660 --> 00:09:06,290
of machine learning who was famous for
221
00:09:04,370 --> 00:09:08,540
building a checklist playing program
222
00:09:06,290 --> 00:09:11,089
that could play checkers even better
223
00:09:08,540 --> 00:09:14,870
than he himself the inventor could play
224
00:09:11,089 --> 00:09:17,769
the game so a machine learning project
225
00:09:14,870 --> 00:09:22,700
will often result in a piece of software
226
00:09:17,769 --> 00:09:26,029
that runs that outputs be given a in
227
00:09:22,700 --> 00:09:28,100
contrast data science is the signs of
228
00:09:26,029 --> 00:09:31,399
extracting knowledge and insights from
229
00:09:28,100 --> 00:09:34,570
data and so the output of a data science
230
00:09:31,399 --> 00:09:37,970
project is often a slide deck Department
231
00:09:34,570 --> 00:09:40,550
presentation that summarizes conclusions
232
00:09:37,970 --> 00:09:42,589
for executives to take business actions
233
00:09:40,550 --> 00:09:45,320
or that summarizes conclusions for a
234
00:09:42,589 --> 00:09:47,959
product team to decide how to improve a
235
00:09:45,320 --> 00:09:50,000
website let me give an example of
236
00:09:47,959 --> 00:09:52,760
machine learning versus data science in
237
00:09:50,000 --> 00:09:54,980
the online advertising industry today
238
00:09:52,760 --> 00:09:57,350
the large high platforms all have a
239
00:09:54,980 --> 00:09:59,000
piece of AI that quickly tells them
240
00:09:57,350 --> 00:10:00,950
what's the ad you are most likely to
241
00:09:59,000 --> 00:10:02,240
click on so that's a machine learning
242
00:10:00,950 --> 00:10:04,370
system and this turns out to be
243
00:10:02,240 --> 00:10:05,930
incredibly lucrative AI system the
244
00:10:04,370 --> 00:10:08,000
inputs information about you and about
245
00:10:05,930 --> 00:10:09,820
the ad and outputs will you click on
246
00:10:08,000 --> 00:10:12,230
this or not these systems are running
247
00:10:09,820 --> 00:10:14,149
24/7 and these are machine learning
248
00:10:12,230 --> 00:10:16,160
systems that drive ad revenue for these
249
00:10:14,149 --> 00:10:18,620
companies so there's a piece of software
250
00:10:16,160 --> 00:10:20,660
that runs in contrast I've also done
251
00:10:18,620 --> 00:10:23,360
data science projects in the online
252
00:10:20,660 --> 00:10:26,510
advertising industry if analyzing data
253
00:10:23,360 --> 00:10:28,760
tells you for example that the travel
254
00:10:26,510 --> 00:10:30,589
industry is not buying a lot of ads but
255
00:10:28,760 --> 00:10:32,779
if you send more sales people to sell
256
00:10:30,589 --> 00:10:34,730
ads the travel companies you could
257
00:10:32,779 --> 00:10:36,980
convince them to use more advertising
258
00:10:34,730 --> 00:10:38,449
then that would be an example of a data
259
00:10:36,980 --> 00:10:39,180
science project and the data science
260
00:10:38,449 --> 00:10:41,579
conclusion
261
00:10:39,180 --> 00:10:43,829
the results in the executives deciding
262
00:10:41,579 --> 00:10:46,529
to ask the sales team to spend more time
263
00:10:43,829 --> 00:10:48,269
reaching out to the travel industry so
264
00:10:46,529 --> 00:10:49,740
even in one company you may have
265
00:10:48,269 --> 00:10:51,629
different machine learning and data
266
00:10:49,740 --> 00:10:55,379
science project spoke for which can be
267
00:10:51,629 --> 00:10:58,050
incredibly valuable you have also heard
268
00:10:55,379 --> 00:11:00,360
of deep learning so what does deep
269
00:10:58,050 --> 00:11:02,910
learning let's say want to predict
270
00:11:00,360 --> 00:11:05,910
housing prices you want to price houses
271
00:11:02,910 --> 00:11:07,319
so you have an input that tells you the
272
00:11:05,910 --> 00:11:08,870
size of house number of bedrooms and
273
00:11:07,319 --> 00:11:11,610
bathrooms and where this newly renovated
274
00:11:08,870 --> 00:11:13,829
one of the most effective ways to priced
275
00:11:11,610 --> 00:11:16,800
houses given this input a will be
276
00:11:13,829 --> 00:11:19,379
defeated this thing here in order to
277
00:11:16,800 --> 00:11:21,389
have it output the price this big thing
278
00:11:19,379 --> 00:11:22,860
in the middle is called a neural network
279
00:11:21,389 --> 00:11:25,740
and sometimes we also call it an
280
00:11:22,860 --> 00:11:27,569
artificial neural network and that's the
281
00:11:25,740 --> 00:11:30,180
distinguish it from the neural network
282
00:11:27,569 --> 00:11:32,939
that is in your brain so the human brain
283
00:11:30,180 --> 00:11:34,850
is made up of neurons and so when we say
284
00:11:32,939 --> 00:11:37,259
artificial neural network that's just
285
00:11:34,850 --> 00:11:39,180
emphasize that this is not the
286
00:11:37,259 --> 00:11:41,879
biological brain but it says a piece of
287
00:11:39,180 --> 00:11:43,410
software and what a neural network does
288
00:11:41,879 --> 00:11:46,439
we're not official neural network does
289
00:11:43,410 --> 00:11:52,249
is take this input a which is all of
290
00:11:46,439 --> 00:11:54,990
these whole things and then output B
291
00:11:52,249 --> 00:11:57,870
which is the estimated price of the
292
00:11:54,990 --> 00:12:00,300
house now in a later optional video this
293
00:11:57,870 --> 00:12:03,449
week I'll show you more what this
294
00:12:00,300 --> 00:12:06,170
artificial neural network really is but
295
00:12:03,449 --> 00:12:07,709
all of human cognition is made up of
296
00:12:06,170 --> 00:12:10,350
neurons in your brain
297
00:12:07,709 --> 00:12:13,050
passing electrical impulses positive low
298
00:12:10,350 --> 00:12:14,790
messages each other and when we draw a
299
00:12:13,050 --> 00:12:16,860
picture of an artificial neural network
300
00:12:14,790 --> 00:12:18,959
there's a very loose analogy to the
301
00:12:16,860 --> 00:12:19,410
brain and these little circles are
302
00:12:18,959 --> 00:12:21,569
called
303
00:12:19,410 --> 00:12:23,970
artificial neurons or just neurons for
304
00:12:21,569 --> 00:12:27,269
short that also passes in neurons to
305
00:12:23,970 --> 00:12:29,129
each other and this big artificial
306
00:12:27,269 --> 00:12:32,009
neural network is just a big
307
00:12:29,129 --> 00:12:34,649
mathematical equation that tells it
308
00:12:32,009 --> 00:12:37,559
given the inputs a how do you compute
309
00:12:34,649 --> 00:12:39,540
the price B in case it seems like there
310
00:12:37,559 --> 00:12:41,639
are a lot of details here don't worry
311
00:12:39,540 --> 00:12:44,220
about it we'll talk more about these
312
00:12:41,639 --> 00:12:46,769
details later but the key takeaways are
313
00:12:44,220 --> 00:12:49,170
that a neural network is a very
314
00:12:46,769 --> 00:12:50,560
effective technique for learning a to be
315
00:12:49,170 --> 00:12:53,140
your input output mapping
316
00:12:50,560 --> 00:12:54,580
and today determines neural network and
317
00:12:53,140 --> 00:12:56,290
deep learning are used almost
318
00:12:54,580 --> 00:12:58,870
interchangeably they mean essentially
319
00:12:56,290 --> 00:13:00,430
the same thing many decades ago this
320
00:12:58,870 --> 00:13:02,320
type of software was called a neural
321
00:13:00,430 --> 00:13:04,300
network but in recent years we found
322
00:13:02,320 --> 00:13:07,180
that you know deep learning was just a
323
00:13:04,300 --> 00:13:08,980
much better sounding brand and so that
324
00:13:07,180 --> 00:13:12,460
thought better versus the term that's
325
00:13:08,980 --> 00:13:14,770
been taking off recently so what do new
326
00:13:12,460 --> 00:13:17,290
networks or artificial neural networks
327
00:13:14,770 --> 00:13:19,480
have to do with the brain it turns out
328
00:13:17,290 --> 00:13:21,730
almost nothing new networks were
329
00:13:19,480 --> 00:13:23,380
originally inspired by the brain but the
330
00:13:21,730 --> 00:13:25,089
details of how they work are almost
331
00:13:23,380 --> 00:13:27,760
completely unrelated to how biological
332
00:13:25,089 --> 00:13:30,420
brains work so I choose very courses
333
00:13:27,760 --> 00:13:32,529
today about making any analogies between
334
00:13:30,420 --> 00:13:34,390
artificial neural networks and the
335
00:13:32,529 --> 00:13:37,420
biological brain even though there was
336
00:13:34,390 --> 00:13:40,300
some loose inspiration there so AI has
337
00:13:37,420 --> 00:13:42,640
many different tools in this video you
338
00:13:40,300 --> 00:13:46,450
learned about what a machine learning
339
00:13:42,640 --> 00:13:49,450
and data science and also what is deep
340
00:13:46,450 --> 00:13:50,830
learning and was it neural network you
341
00:13:49,450 --> 00:13:52,600
might also hear in the media other
342
00:13:50,830 --> 00:13:54,310
buzzwords like unsupervised learning
343
00:13:52,600 --> 00:13:56,080
wrinkles learning graphic novels
344
00:13:54,310 --> 00:13:58,030
planning knowledge drop and so on and
345
00:13:56,080 --> 00:13:59,470
you don't need to know what all of these
346
00:13:58,030 --> 00:14:02,050
other terms mean but these are just
347
00:13:59,470 --> 00:14:04,390
other tools for getting AI systems to
348
00:14:02,050 --> 00:14:05,800
make computers act intelligent you know
349
00:14:04,390 --> 00:14:08,020
try to give you a sense of what some of
350
00:14:05,800 --> 00:14:11,290
these terms mean in later videos as well
351
00:14:08,020 --> 00:14:13,960
but the most important tools that I hope
352
00:14:11,290 --> 00:14:16,089
you know about are machine learning and
353
00:14:13,960 --> 00:14:17,560
data science as well as deep learning
354
00:14:16,089 --> 00:14:20,140
the neural networks which are a very
355
00:14:17,560 --> 00:14:22,930
powerful way to do machine learning and
356
00:14:20,140 --> 00:14:24,670
sometimes data science if we were to
357
00:14:22,930 --> 00:14:26,290
draw a Venn diagram showing how all
358
00:14:24,670 --> 00:14:30,130
these concepts fit together and this is
359
00:14:26,290 --> 00:14:32,890
what it might look like AI is this huge
360
00:14:30,130 --> 00:14:36,730
set of tools for making computers behave
361
00:14:32,890 --> 00:14:40,240
intelligently of AI the biggest subset
362
00:14:36,730 --> 00:14:42,880
is very tools from machine learning but
363
00:14:40,240 --> 00:14:44,380
AI does have other tools than machine
364
00:14:42,880 --> 00:14:47,650
learning such as some of these buzz
365
00:14:44,380 --> 00:14:49,600
words are listed at the bottom and of
366
00:14:47,650 --> 00:14:51,250
machine learning the part of machine
367
00:14:49,600 --> 00:14:52,000
learning that's most important these
368
00:14:51,250 --> 00:14:54,820
days is
369
00:14:52,000 --> 00:14:56,800
neural networks or deep learning which
370
00:14:54,820 --> 00:14:58,959
is very powerful set of tools for
371
00:14:56,800 --> 00:15:00,730
carrying out supervised learning or a to
372
00:14:58,959 --> 00:15:02,890
be mappings as well as some other things
373
00:15:00,730 --> 00:15:03,160
but they're also other machine learnings
374
00:15:02,890 --> 00:15:07,060
who
375
00:15:03,160 --> 00:15:08,649
that are not just deep learning tools so
376
00:15:07,060 --> 00:15:10,930
how does data science fit into this
377
00:15:08,649 --> 00:15:12,940
picture there is inconsistency in how
378
00:15:10,930 --> 00:15:15,009
two terminologies use some people will
379
00:15:12,940 --> 00:15:17,230
tell you the designs is a subset of AI
380
00:15:15,009 --> 00:15:18,699
some people will tell you AI is a subset
381
00:15:17,230 --> 00:15:19,870
that they design so it depends a bit on
382
00:15:18,699 --> 00:15:21,550
who you ask
383
00:15:19,870 --> 00:15:25,660
but I would say that data science is
384
00:15:21,550 --> 00:15:29,500
maybe a cross-cutting subset of all of
385
00:15:25,660 --> 00:15:30,970
these tools that uses many tools from AI
386
00:15:29,500 --> 00:15:32,889
machine learning and deep learning but
387
00:15:30,970 --> 00:15:35,319
has some other separate tools as well
388
00:15:32,889 --> 00:15:38,649
that solves a very set of important
389
00:15:35,319 --> 00:15:40,660
problems in driving business insights in
390
00:15:38,649 --> 00:15:42,610
this video you saw what is machine
391
00:15:40,660 --> 00:15:44,920
learning where the state designs and
392
00:15:42,610 --> 00:15:46,689
what is deep learning and neural
393
00:15:44,920 --> 00:15:48,670
networks I hope this gives you a sense
394
00:15:46,689 --> 00:15:51,129
of the most common and important
395
00:15:48,670 --> 00:15:52,660
terminology using AI and you can start
396
00:15:51,129 --> 00:15:55,959
thinking about how these things might
397
00:15:52,660 --> 00:15:58,410
apply to your company now what does it
398
00:15:55,959 --> 00:16:02,970
mean for a company to be good at AI
399
00:15:58,410 --> 00:16:02,970
let's talk about that in the next video
400
00:16:08,400 --> 00:16:10,460
you
401
00:17:32,160 --> 00:17:37,290
what makes a company good at AI and
402
00:17:34,770 --> 00:17:39,690
perhaps even more importantly what would
403
00:17:37,290 --> 00:17:42,960
it take for your country to become great
404
00:17:39,690 --> 00:17:45,360
and using AI I had previously led the
405
00:17:42,960 --> 00:17:47,760
Google brain team and by deuce AI group
406
00:17:45,360 --> 00:17:51,480
which I respectively helped Google and
407
00:17:47,760 --> 00:17:53,400
Baidu become great AI companies so what
408
00:17:51,480 --> 00:17:55,890
can you do for your company
409
00:17:53,400 --> 00:17:57,840
thus a lesson had learned washing the
410
00:17:55,890 --> 00:17:59,970
rise of the internet that I think would
411
00:17:57,840 --> 00:18:02,690
be relevant to how all of us navigate
412
00:17:59,970 --> 00:18:04,860
the rise of AI let's take a look a
413
00:18:02,690 --> 00:18:06,660
lesson we learned from the rise of the
414
00:18:04,860 --> 00:18:09,030
internet was that if you take your
415
00:18:06,660 --> 00:18:10,860
favorite shopping mall so you know my
416
00:18:09,030 --> 00:18:12,630
wife and I sometimes shop at Stanford
417
00:18:10,860 --> 00:18:14,280
Shopping Center and you put a website
418
00:18:12,630 --> 00:18:17,100
for the shopping mall maybe sell things
419
00:18:14,280 --> 00:18:19,140
on the website that by itself does not
420
00:18:17,100 --> 00:18:22,050
turn the shopping mall into an Internet
421
00:18:19,140 --> 00:18:24,750
company in fact a few years ago I was
422
00:18:22,050 --> 00:18:27,690
speaking with the CEO of a large retail
423
00:18:24,750 --> 00:18:29,370
company who said to me hey Andrew I have
424
00:18:27,690 --> 00:18:31,830
a website I sell things on the website
425
00:18:29,370 --> 00:18:34,080
amazon has a website Amazon sells things
426
00:18:31,830 --> 00:18:35,670
a website is the same thing but of
427
00:18:34,080 --> 00:18:37,920
course it wasn't in the shopping mall
428
00:18:35,670 --> 00:18:41,190
with a website isn't the same thing as a
429
00:18:37,920 --> 00:18:43,170
first-class Internet company so what is
430
00:18:41,190 --> 00:18:44,910
it that defines an Internet company if
431
00:18:43,170 --> 00:18:46,890
it isn't just whether or not you sell
432
00:18:44,910 --> 00:18:49,770
things on the website I think an
433
00:18:46,890 --> 00:18:51,450
Internet company is a company that does
434
00:18:49,770 --> 00:18:53,880
the things the internet lets you do
435
00:18:51,450 --> 00:18:56,100
really well for example we engage in
436
00:18:53,880 --> 00:18:58,260
pervasive AP testing meaning we
437
00:18:56,100 --> 00:19:00,090
routinely throw up two different
438
00:18:58,260 --> 00:19:01,860
versions of web site and see which one
439
00:19:00,090 --> 00:19:03,780
works better because we can and so we
440
00:19:01,860 --> 00:19:05,580
learn much faster whereas in a
441
00:19:03,780 --> 00:19:07,740
traditional shopping mall you know very
442
00:19:05,580 --> 00:19:09,750
difficult to have two shopping malls in
443
00:19:07,740 --> 00:19:12,360
two parallel universes and you can only
444
00:19:09,750 --> 00:19:15,060
maybe chase things around every quarter
445
00:19:12,360 --> 00:19:16,800
every six months Internet companies tend
446
00:19:15,060 --> 00:19:18,570
to have very short iteration time so you
447
00:19:16,800 --> 00:19:20,580
can ship a new product every week or
448
00:19:18,570 --> 00:19:22,290
maybe even every day because you can
449
00:19:20,580 --> 00:19:24,360
wear as a shopping mall can be
450
00:19:22,290 --> 00:19:27,660
redesigned and we are protected only
451
00:19:24,360 --> 00:19:30,600
every several months Internet companies
452
00:19:27,660 --> 00:19:32,790
also tend to push decision making down
453
00:19:30,600 --> 00:19:34,320
from the CEO to the engineers and to
454
00:19:32,790 --> 00:19:35,520
other specialized roles such as the
455
00:19:34,320 --> 00:19:37,679
product managers
456
00:19:35,520 --> 00:19:39,570
this is in contrast to a traditional
457
00:19:37,679 --> 00:19:41,850
shopping mall where you can maybe have
458
00:19:39,570 --> 00:19:43,529
the CEO just decide all the key
459
00:19:41,850 --> 00:19:45,929
decisions and then just everyone does
460
00:19:43,529 --> 00:19:48,149
what the CEO says and it turns out that
461
00:19:45,929 --> 00:19:50,460
traditional model doesn't work in the
462
00:19:48,149 --> 00:19:52,440
internet error because only the
463
00:19:50,460 --> 00:19:54,510
engineers and other specialized roles
464
00:19:52,440 --> 00:19:56,760
like product managers know enough about
465
00:19:54,510 --> 00:19:59,610
the technology and the product and the
466
00:19:56,760 --> 00:20:01,500
users to make great decisions so these
467
00:19:59,610 --> 00:20:04,200
are some of the things that internet
468
00:20:01,500 --> 00:20:05,789
companies do in order to make sure they
469
00:20:04,200 --> 00:20:09,330
do the things that the internet doesn't
470
00:20:05,789 --> 00:20:12,360
do really well this is a lesson we learn
471
00:20:09,330 --> 00:20:12,899
from the internet error how about the AI
472
00:20:12,360 --> 00:20:16,049
error
473
00:20:12,899 --> 00:20:18,750
I think that today you can take any
474
00:20:16,049 --> 00:20:20,429
company and haven't used a few neural
475
00:20:18,750 --> 00:20:23,100
networks or few deep learning algorithms
476
00:20:20,429 --> 00:20:26,190
that by itself does not turn the company
477
00:20:23,100 --> 00:20:28,049
into an AI company instead what makes a
478
00:20:26,190 --> 00:20:31,110
great AI company is sometimes an AI
479
00:20:28,049 --> 00:20:33,120
first company is are you doing the
480
00:20:31,110 --> 00:20:36,240
things that AI lets you do really well
481
00:20:33,120 --> 00:20:38,880
for example AI companies are very good
482
00:20:36,240 --> 00:20:41,190
at strategic data acquisition this is
483
00:20:38,880 --> 00:20:43,679
why many of the large consumer tech
484
00:20:41,190 --> 00:20:46,289
companies may have three products that
485
00:20:43,679 --> 00:20:48,029
do not monetize and it allows them to
486
00:20:46,289 --> 00:20:51,419
acquire data that they can monetize
487
00:20:48,029 --> 00:20:53,549
elsewhere so let strategy teams where we
488
00:20:51,419 --> 00:20:55,740
would deliberately launch products that
489
00:20:53,549 --> 00:20:58,799
do not make any money just for the sake
490
00:20:55,740 --> 00:21:02,460
of data acquisition and thinking through
491
00:20:58,799 --> 00:21:05,730
how to get data is a key part of the
492
00:21:02,460 --> 00:21:09,120
great AI companies a company sends up
493
00:21:05,730 --> 00:21:11,309
unified data warehouses if you have 50
494
00:21:09,120 --> 00:21:13,710
different databases or 50 different data
495
00:21:11,309 --> 00:21:15,809
warehouses under the control of 50
496
00:21:13,710 --> 00:21:18,750
different vice-presidents then they'll
497
00:21:15,809 --> 00:21:20,520
be impossible for an engineer to get the
498
00:21:18,750 --> 00:21:22,500
data into one place so that they can
499
00:21:20,520 --> 00:21:24,380
connect the dots and swap the patterns
500
00:21:22,500 --> 00:21:27,090
so many great our companies have
501
00:21:24,380 --> 00:21:29,549
preemptively invested in bringing the
502
00:21:27,090 --> 00:21:31,740
data together into a single data
503
00:21:29,549 --> 00:21:33,450
warehouse to increase the odds that the
504
00:21:31,740 --> 00:21:36,000
teams can connect the dots
505
00:21:33,450 --> 00:21:39,600
subjective cause to privacy guarantees
506
00:21:36,000 --> 00:21:41,909
and also to data regulations such as GDP
507
00:21:39,600 --> 00:21:43,380
are in Europe our companies are very
508
00:21:41,909 --> 00:21:45,149
good at spotting automation
509
00:21:43,380 --> 00:21:47,159
opportunities we're very good at seeing
510
00:21:45,149 --> 00:21:47,950
oh let's insert the supervised learning
511
00:21:47,159 --> 00:21:50,620
Albert
512
00:21:47,950 --> 00:21:52,450
and have a a to be mapping here so that
513
00:21:50,620 --> 00:21:55,120
we don't have to have people do these
514
00:21:52,450 --> 00:21:57,580
tasks instead we can automate it yeah I
515
00:21:55,120 --> 00:22:00,549
companies also have many new roles such
516
00:21:57,580 --> 00:22:03,610
as the MLE or machine learning engineer
517
00:22:00,549 --> 00:22:06,610
and new ways of dividing up toss among
518
00:22:03,610 --> 00:22:08,760
different members of the team so for
519
00:22:06,610 --> 00:22:11,230
company to become good at AI means
520
00:22:08,760 --> 00:22:13,750
architecting the company to do the
521
00:22:11,230 --> 00:22:16,389
things that AI makes it possible to do
522
00:22:13,750 --> 00:22:19,690
really well now for a company to become
523
00:22:16,389 --> 00:22:21,970
good at AI does require a process in
524
00:22:19,690 --> 00:22:23,200
fact 10 years ago Google and Baidu as
525
00:22:21,970 --> 00:22:25,090
well as companies like Facebook and
526
00:22:23,200 --> 00:22:27,010
Microsoft that was now the part of we're
527
00:22:25,090 --> 00:22:30,190
not great AI company is the way that
528
00:22:27,010 --> 00:22:33,669
they are today so how can a company
529
00:22:30,190 --> 00:22:36,340
become good at AI it turns out that
530
00:22:33,669 --> 00:22:38,230
becoming good at AI is not a mysterious
531
00:22:36,340 --> 00:22:40,600
magical process instead there is a
532
00:22:38,230 --> 00:22:42,909
systematic process through which many
533
00:22:40,600 --> 00:22:46,120
companies almost any big company can
534
00:22:42,909 --> 00:22:48,429
become good at AI this is the five-step
535
00:22:46,120 --> 00:22:50,110
AI transformation playbook that I
536
00:22:48,429 --> 00:22:53,080
recommend to companies that want to
537
00:22:50,110 --> 00:22:54,820
become effective and using AI I'll give
538
00:22:53,080 --> 00:22:57,010
a brief overview of the playbook here
539
00:22:54,820 --> 00:23:00,340
and then go into detail in a later week
540
00:22:57,010 --> 00:23:02,409
step one is to execute pilot projects to
541
00:23:00,340 --> 00:23:04,600
gain momentum so just do a few small
542
00:23:02,409 --> 00:23:06,580
projects to get Ben a sense of what AI
543
00:23:04,600 --> 00:23:09,279
can and cannot do and get a better sense
544
00:23:06,580 --> 00:23:11,110
of what doing an AI project feels like
545
00:23:09,279 --> 00:23:13,779
and this you could do in-house or you
546
00:23:11,110 --> 00:23:15,639
can also do with an outsource team but
547
00:23:13,779 --> 00:23:17,830
eventually you then need to do step 2
548
00:23:15,639 --> 00:23:21,669
which is to build an in-house AI team
549
00:23:17,830 --> 00:23:23,500
and provide broad AI training not just
550
00:23:21,669 --> 00:23:25,419
to the engineers but also to the
551
00:23:23,500 --> 00:23:28,179
managers division leaders and executives
552
00:23:25,419 --> 00:23:29,860
and how to think about AI after doing
553
00:23:28,179 --> 00:23:33,130
this so as you're doing this you have a
554
00:23:29,860 --> 00:23:34,990
better sense of what AI is and then it's
555
00:23:33,130 --> 00:23:38,860
important for many companies to develop
556
00:23:34,990 --> 00:23:41,769
an AI strategy and finally to align
557
00:23:38,860 --> 00:23:43,269
internal and external communications so
558
00:23:41,769 --> 00:23:45,370
that all your stakeholders from
559
00:23:43,269 --> 00:23:47,580
employees customers and investors are
560
00:23:45,370 --> 00:23:50,799
aligned with how your company is
561
00:23:47,580 --> 00:23:53,169
navigating the rise of AI the AI has
562
00:23:50,799 --> 00:23:55,779
created tremendous value in the software
563
00:23:53,169 --> 00:23:57,370
industry and will continue to do so it
564
00:23:55,779 --> 00:23:59,240
will also create tremendous value
565
00:23:57,370 --> 00:24:01,429
outside the software industry
566
00:23:59,240 --> 00:24:03,500
if you can help your company become good
567
00:24:01,429 --> 00:24:06,890
at AI I hope you can play a leading role
568
00:24:03,500 --> 00:24:09,049
in creating a lot of this value in this
569
00:24:06,890 --> 00:24:11,570
video you saw what is it that makes a
570
00:24:09,049 --> 00:24:13,730
company a good AI company and also
571
00:24:11,570 --> 00:24:15,320
briefly the AI transformation playbook
572
00:24:13,730 --> 00:24:18,049
which they're going to much create a
573
00:24:15,320 --> 00:24:20,299
detail on in a later week as a roadmap
574
00:24:18,049 --> 00:24:22,580
for helping companies become great at AI
575
00:24:20,299 --> 00:24:24,980
if you're interested there is also
576
00:24:22,580 --> 00:24:26,929
published online an AI transformation
577
00:24:24,980 --> 00:24:28,820
playbook that goes into these five steps
578
00:24:26,929 --> 00:24:31,669
in greater detail but you see more of
579
00:24:28,820 --> 00:24:34,100
these in the later league as well now
580
00:24:31,669 --> 00:24:35,960
one of the challenges of doing a are
581
00:24:34,100 --> 00:24:38,649
projects such as the pilot project in
582
00:24:35,960 --> 00:24:38,649
step one is
583
00:24:40,240 --> 00:24:45,820
in this video and the next video I hope
584
00:24:43,480 --> 00:24:49,419
to help you develop intuition about what
585
00:24:45,820 --> 00:24:51,490
a I can and cannot do in practice before
586
00:24:49,419 --> 00:24:53,950
I commit to a specific AI project I'll
587
00:24:51,490 --> 00:24:56,529
usually have either myself or engineers
588
00:24:53,950 --> 00:24:58,419
do technical diligence on the project to
589
00:24:56,529 --> 00:25:00,610
make sure that it is feasible this means
590
00:24:58,419 --> 00:25:02,649
look in the data look at the input and
591
00:25:00,610 --> 00:25:04,990
output a and B and just thinking through
592
00:25:02,649 --> 00:25:07,149
if this is something a I can really do
593
00:25:04,990 --> 00:25:09,690
what I've seen unfortunately is that
594
00:25:07,149 --> 00:25:12,249
some CEOs can have an overinflated
595
00:25:09,690 --> 00:25:14,649
expectation of AI and can ask engineers
596
00:25:12,249 --> 00:25:17,169
to do things that today's AI just cannot
597
00:25:14,649 --> 00:25:18,940
do one of the challenges is that the
598
00:25:17,169 --> 00:25:22,179
media as well as the academic literature
599
00:25:18,940 --> 00:25:24,519
tends to only report on positive results
600
00:25:22,179 --> 00:25:26,889
of success stories using AI and we see a
601
00:25:24,519 --> 00:25:29,169
string of success stories and no failure
602
00:25:26,889 --> 00:25:31,389
stories people sometimes think AI can do
603
00:25:29,169 --> 00:25:33,909
everything and unfortunately that's just
604
00:25:31,389 --> 00:25:35,919
not true so what I want to do in this
605
00:25:33,909 --> 00:25:37,840
and the next video is to show you a few
606
00:25:35,919 --> 00:25:41,049
examples of what today's AI technology
607
00:25:37,840 --> 00:25:42,460
can do but also what it cannot do and I
608
00:25:41,049 --> 00:25:44,799
hope that this will help you hone your
609
00:25:42,460 --> 00:25:47,289
intuition about what might be more or
610
00:25:44,799 --> 00:25:49,929
less promising projects to select for
611
00:25:47,289 --> 00:25:52,779
your company previously you saw this
612
00:25:49,929 --> 00:25:53,919
list of AI applications from span 14 to
613
00:25:52,779 --> 00:25:57,549
speech recognition to machine
614
00:25:53,919 --> 00:25:59,649
translation and so on one imperfect rule
615
00:25:57,549 --> 00:26:01,269
of thumb you can use to decide what
616
00:25:59,649 --> 00:26:03,429
supervised learning may or may not be
617
00:26:01,269 --> 00:26:05,860
able to do is that pretty much anything
618
00:26:03,429 --> 00:26:08,559
you can do with a second of thought we
619
00:26:05,860 --> 00:26:10,259
can probably now or soon automates using
620
00:26:08,559 --> 00:26:14,139
supervised learning using this
621
00:26:10,259 --> 00:26:16,749
input/output mapping so for example in
622
00:26:14,139 --> 00:26:18,759
order to determine the position of other
623
00:26:16,749 --> 00:26:21,639
costs you know that's something that you
624
00:26:18,759 --> 00:26:25,539
can do with less than a second in order
625
00:26:21,639 --> 00:26:27,070
to tell if a phone is strache you can
626
00:26:25,539 --> 00:26:29,259
look at it and you can kind of tell in
627
00:26:27,070 --> 00:26:31,869
less than a second in order to
628
00:26:29,259 --> 00:26:33,460
understand at least transcribe what was
629
00:26:31,869 --> 00:26:35,799
said you know doesn't take that many
630
00:26:33,460 --> 00:26:38,350
seconds of thought and while this is an
631
00:26:35,799 --> 00:26:40,899
imperfect rule of thumb it maybe gives
632
00:26:38,350 --> 00:26:44,200
you a way to quickly think of some
633
00:26:40,899 --> 00:26:46,600
examples of tasks that AI systems can do
634
00:26:44,200 --> 00:26:49,450
whereas in contrast something that AI
635
00:26:46,600 --> 00:26:50,759
today cannot do would be to analyze a
636
00:26:49,450 --> 00:26:52,889
market and write
637
00:26:50,759 --> 00:26:55,049
50 page report the human cannot write a
638
00:26:52,889 --> 00:26:57,449
50 page market analysis report in a
639
00:26:55,049 --> 00:26:59,549
second and it's very difficult and he's
640
00:26:57,449 --> 00:27:01,949
I don't know and I don't think any team
641
00:26:59,549 --> 00:27:04,529
in the world today knows how to get an
642
00:27:01,949 --> 00:27:07,529
AI system to do market research and run
643
00:27:04,529 --> 00:27:09,139
an extended market report either I found
644
00:27:07,529 --> 00:27:12,509
that one of the best ways the whole
645
00:27:09,139 --> 00:27:14,519
intuition is a look at concrete examples
646
00:27:12,509 --> 00:27:17,190
so let's take a look at a specific
647
00:27:14,519 --> 00:27:19,349
example relating to customer support
648
00:27:17,190 --> 00:27:21,119
automation let's see you run a website
649
00:27:19,349 --> 00:27:23,309
there sell things so an e-commerce
650
00:27:21,119 --> 00:27:25,139
company and you have a Customer Support
651
00:27:23,309 --> 00:27:27,209
Division that gets an email like this
652
00:27:25,139 --> 00:27:28,199
the tour arrived two days later I wasn't
653
00:27:27,209 --> 00:27:31,199
going to give it to my niece for her
654
00:27:28,199 --> 00:27:33,359
birthday can I return it if what you
655
00:27:31,199 --> 00:27:36,479
want is an AI system that looks at this
656
00:27:33,359 --> 00:27:38,369
and decides this is a refund request
657
00:27:36,479 --> 00:27:40,319
so let me route it to my refund
658
00:27:38,369 --> 00:27:42,359
department then I was saying you have a
659
00:27:40,319 --> 00:27:45,690
good chance of building an AI system to
660
00:27:42,359 --> 00:27:48,809
do that the AI system would take as
661
00:27:45,690 --> 00:27:51,569
input the customer checks what the
662
00:27:48,809 --> 00:27:54,359
customer emails you and it would output
663
00:27:51,569 --> 00:27:56,209
this is a refund request or is this a
664
00:27:54,359 --> 00:27:59,609
shipping problem or is it a other
665
00:27:56,209 --> 00:28:01,169
requests in order to route this email to
666
00:27:59,609 --> 00:28:04,049
the most appropriate part of your
667
00:28:01,169 --> 00:28:07,529
customer support center so the input aid
668
00:28:04,049 --> 00:28:09,059
is at X and the output B is one of these
669
00:28:07,529 --> 00:28:11,099
three outcomes there's a refund or a
670
00:28:09,059 --> 00:28:13,440
shipping problem or shipping query or is
671
00:28:11,099 --> 00:28:16,079
it a different request so this is
672
00:28:13,440 --> 00:28:18,659
something that AI today can do here's
673
00:28:16,079 --> 00:28:21,239
something a act today cannot do which is
674
00:28:18,659 --> 00:28:23,729
if you want AI to input an email and
675
00:28:21,239 --> 00:28:25,559
automatically generate a response like
676
00:28:23,729 --> 00:28:27,209
oh sorry here that I hope you needed a
677
00:28:25,559 --> 00:28:27,719
good birthday yes we can help work and
678
00:28:27,209 --> 00:28:30,719
so on
679
00:28:27,719 --> 00:28:32,940
so for an AI to output a complicated
680
00:28:30,719 --> 00:28:35,519
piece of text like this today is very
681
00:28:32,940 --> 00:28:37,739
difficult by today's standards of AI and
682
00:28:35,519 --> 00:28:40,109
in fact to even empathize about the
683
00:28:37,739 --> 00:28:42,299
birthday of your knees that is very
684
00:28:40,109 --> 00:28:44,729
difficult to do for every single
685
00:28:42,299 --> 00:28:47,639
possible type of email you might receive
686
00:28:44,729 --> 00:28:49,769
now what would happen if you were to use
687
00:28:47,639 --> 00:28:51,599
a machine learning tool like a deep
688
00:28:49,769 --> 00:28:53,999
learning algorithm to try to do this
689
00:28:51,599 --> 00:28:56,459
anyway so let's say you try to get an AI
690
00:28:53,999 --> 00:28:58,829
system to input the user's email and
691
00:28:56,459 --> 00:29:01,139
open a to the three-paragraph
692
00:28:58,829 --> 00:29:03,119
empathetic and appropriate response and
693
00:29:01,139 --> 00:29:04,169
let's say that you have a mother size
694
00:29:03,119 --> 00:29:06,899
data set you're like
695
00:29:04,169 --> 00:29:09,509
thousand examples of user emails and
696
00:29:06,899 --> 00:29:11,639
appropriate responses it turns out if
697
00:29:09,509 --> 00:29:13,859
you run an AI system on this type of
698
00:29:11,639 --> 00:29:16,139
data on a small data set like a thousand
699
00:29:13,859 --> 00:29:18,989
examples this may be the performance you
700
00:29:16,139 --> 00:29:20,730
get which is if a user emails my box is
701
00:29:18,989 --> 00:29:22,950
damaged you'll say thank you for email
702
00:29:20,730 --> 00:29:25,499
and it says whether rather review thank
703
00:29:22,950 --> 00:29:27,929
you email what's written policy thank
704
00:29:25,499 --> 00:29:29,879
you very much but the problem we're
705
00:29:27,929 --> 00:29:32,159
building this type of AI is that with
706
00:29:29,879 --> 00:29:34,350
just a thousand examples there's just
707
00:29:32,159 --> 00:29:35,999
not enough data for an AI system to
708
00:29:34,350 --> 00:29:38,009
learn how to write to the three
709
00:29:35,999 --> 00:29:40,559
paragraph appropriate and empathetic
710
00:29:38,009 --> 00:29:42,929
responses so it may end up just
711
00:29:40,559 --> 00:29:45,299
generating the same very simple response
712
00:29:42,929 --> 00:29:48,539
like thank you very mad what the
713
00:29:45,299 --> 00:29:50,070
customer is sending you another thing
714
00:29:48,539 --> 00:29:51,960
that could go wrong another way from the
715
00:29:50,070 --> 00:29:54,090
assistance of fail is if it generates
716
00:29:51,960 --> 00:29:56,340
gibberish such as whether my boss
717
00:29:54,090 --> 00:29:58,919
arriving and it says thank yes now
718
00:29:56,340 --> 00:30:01,049
you're kind of gibberish and this is a
719
00:29:58,919 --> 00:30:03,029
hard enough problem that even with ten
720
00:30:01,049 --> 00:30:05,399
thousand or a hundred thousand email
721
00:30:03,029 --> 00:30:07,919
examples I don't know if that would be
722
00:30:05,399 --> 00:30:10,859
enough data for an AI system to do this
723
00:30:07,919 --> 00:30:13,980
well the rules for what AI can it cannot
724
00:30:10,859 --> 00:30:16,259
do are not hard and fast and I usually
725
00:30:13,980 --> 00:30:18,239
end up having to ask engineering teams
726
00:30:16,259 --> 00:30:20,999
to sometimes spend a few weeks doing
727
00:30:18,239 --> 00:30:24,149
deep technical diligence to decide for
728
00:30:20,999 --> 00:30:26,549
myself if a project is feasible but to
729
00:30:24,149 --> 00:30:28,259
hone your intuitions to help you quickly
730
00:30:26,549 --> 00:30:29,970
filter feasible and not feasible
731
00:30:28,259 --> 00:30:32,190
projects here are a couple of other
732
00:30:29,970 --> 00:30:34,649
rules of thumb about what makes a
733
00:30:32,190 --> 00:30:37,889
machine learning problem easier or more
734
00:30:34,649 --> 00:30:40,919
likely to be feasible one learning a
735
00:30:37,889 --> 00:30:43,889
simple concept is more likely to be
736
00:30:40,919 --> 00:30:45,840
feasible well what is a simple concept
737
00:30:43,889 --> 00:30:47,850
mean there's no formal definition of
738
00:30:45,840 --> 00:30:50,369
that but this is something that takes
739
00:30:47,850 --> 00:30:52,019
you less than a second of mental thought
740
00:30:50,369 --> 00:30:53,580
or a very very small number of seconds
741
00:30:52,019 --> 00:30:55,799
of mental thought to come up with a
742
00:30:53,580 --> 00:30:58,230
conclusion then that would lean to
743
00:30:55,799 --> 00:31:00,059
whether it being a simple concept so
744
00:30:58,230 --> 00:31:01,980
you're looking outside the window of a
745
00:31:00,059 --> 00:31:03,840
self-driving car to spot the other calls
746
00:31:01,980 --> 00:31:04,830
that would be a relatively simple
747
00:31:03,840 --> 00:31:06,509
concept
748
00:31:04,830 --> 00:31:08,309
whereas how to write an empathetic
749
00:31:06,509 --> 00:31:11,279
response to a complicated user
750
00:31:08,309 --> 00:31:12,730
complaints that would be less of a
751
00:31:11,279 --> 00:31:14,650
simple concept
752
00:31:12,730 --> 00:31:16,630
seconds a machine learning problem is
753
00:31:14,650 --> 00:31:19,480
more likely to be feasible if you have
754
00:31:16,630 --> 00:31:23,260
lots of data available and here our data
755
00:31:19,480 --> 00:31:26,679
means both the input a and the output B
756
00:31:23,260 --> 00:31:30,730
that you want the AI system to have in
757
00:31:26,679 --> 00:31:32,410
your a to be input-output mapping so for
758
00:31:30,730 --> 00:31:35,110
example in the customer support
759
00:31:32,410 --> 00:31:38,320
application the input a would be
760
00:31:35,110 --> 00:31:40,809
examples of emails from customers and B
761
00:31:38,320 --> 00:31:43,840
could be labeling each of these custom
762
00:31:40,809 --> 00:31:46,540
emails as to whether is a refund request
763
00:31:43,840 --> 00:31:48,520
or a shipping query or some other
764
00:31:46,540 --> 00:31:50,919
problem one of the outcomes that if you
765
00:31:48,520 --> 00:31:52,960
have thousands of emails with both a and
766
00:31:50,919 --> 00:31:54,429
B then the odds of you being with build
767
00:31:52,960 --> 00:31:57,580
a machine learning system to do that
768
00:31:54,429 --> 00:31:59,650
would be pretty good AI is the new
769
00:31:57,580 --> 00:32:02,110
electricity and is transforming every
770
00:31:59,650 --> 00:32:04,120
industry but there's also not magic and
771
00:32:02,110 --> 00:32:06,549
you can't do everything under the Sun I
772
00:32:04,120 --> 00:32:08,559
hope that this video started to help you
773
00:32:06,549 --> 00:32:10,720
hone your intuitions about what it can
774
00:32:08,559 --> 00:32:13,150
and cannot do and increase the odds of
775
00:32:10,720 --> 00:32:15,340
your selecting feasible and valuable
776
00:32:13,150 --> 00:32:18,280
projects for maybe your teams to try
777
00:32:15,340 --> 00:32:20,350
working on in order to help you continue
778
00:32:18,280 --> 00:32:23,200
developing your intuition I would like
779
00:32:20,350 --> 00:32:23,919
to show you more examples of what AI can
780
00:32:23,200 --> 00:32:27,330
and cannot do
781
00:32:23,919 --> 00:32:27,330
let's go into the next video
782
00:32:29,799 --> 00:32:35,019
one of the challenges of becoming good
783
00:32:32,379 --> 00:32:37,360
at recognizing what AI can and cannot do
784
00:32:35,019 --> 00:32:39,820
is that it does take seeing a few
785
00:32:37,360 --> 00:32:42,970
examples of concrete successes and
786
00:32:39,820 --> 00:32:45,340
failures of AI and if you work on an
787
00:32:42,970 --> 00:32:47,590
average of say one nu AI project a year
788
00:32:45,340 --> 00:32:49,749
then to see three examples would take
789
00:32:47,590 --> 00:32:51,999
you the years of work experience and
790
00:32:49,749 --> 00:32:54,369
that's just a long time whether how to
791
00:32:51,999 --> 00:32:56,830
do both in the previous video and in
792
00:32:54,369 --> 00:32:59,289
this video is to quickly show you a few
793
00:32:56,830 --> 00:33:01,450
examples of AI successes invidious or
794
00:32:59,289 --> 00:33:03,609
what it can and cannot do so that in a
795
00:33:01,450 --> 00:33:05,859
much shorter time you can see multiple
796
00:33:03,609 --> 00:33:08,379
concrete examples to help hone your
797
00:33:05,859 --> 00:33:10,570
intuition and select valuable projects
798
00:33:08,379 --> 00:33:12,700
so let's take a look at a few more
799
00:33:10,570 --> 00:33:14,649
examples let's say you're building a
800
00:33:12,700 --> 00:33:17,019
self-driving car here's something that
801
00:33:14,649 --> 00:33:18,970
AI can do pretty well which is to take a
802
00:33:17,019 --> 00:33:21,970
picture of what's in front of your car
803
00:33:18,970 --> 00:33:23,950
and maybe just using camera may be using
804
00:33:21,970 --> 00:33:27,730
other senses as well such as radar or
805
00:33:23,950 --> 00:33:30,429
lidar and then to figure out what is the
806
00:33:27,730 --> 00:33:33,340
position or where are the other costs so
807
00:33:30,429 --> 00:33:35,499
this would be an AI where the input a is
808
00:33:33,340 --> 00:33:37,899
a picture of what's in front of your car
809
00:33:35,499 --> 00:33:40,299
or maybe both a picture as well as radar
810
00:33:37,899 --> 00:33:43,809
and other sensor readings and the output
811
00:33:40,299 --> 00:33:45,549
B is where are the other costs and today
812
00:33:43,809 --> 00:33:46,899
the self-driving car industry has
813
00:33:45,549 --> 00:33:49,210
figured out how to collect enough data
814
00:33:46,899 --> 00:33:51,879
and has pretty good algorithms so doing
815
00:33:49,210 --> 00:33:54,580
this reasonably well so that's what a AI
816
00:33:51,879 --> 00:33:57,159
today can do here's an example of
817
00:33:54,580 --> 00:33:58,480
something that today's AI cannot do and
818
00:33:57,159 --> 00:34:00,909
this would be very difficult using
819
00:33:58,480 --> 00:34:03,789
today's AI which is to input a picture
820
00:34:00,909 --> 00:34:05,769
and outputs the intention or whatever
821
00:34:03,789 --> 00:34:07,269
the human is trying to gesture at your
822
00:34:05,769 --> 00:34:09,940
car so here's a construction worker
823
00:34:07,269 --> 00:34:12,280
holding out a hand to ask your car to
824
00:34:09,940 --> 00:34:14,980
stop here's a hitchhiker trying to wave
825
00:34:12,280 --> 00:34:16,569
a car over here's a bicyclist raising
826
00:34:14,980 --> 00:34:19,450
the left hand to indicate that they want
827
00:34:16,569 --> 00:34:21,700
to turn left and so if you were to try
828
00:34:19,450 --> 00:34:24,010
to build a system to learn an ADA be
829
00:34:21,700 --> 00:34:26,230
mapping where the input a is a short
830
00:34:24,010 --> 00:34:28,569
video of a human gesturing at your car
831
00:34:26,230 --> 00:34:31,780
and the output B is what's the intention
832
00:34:28,569 --> 00:34:34,869
what does this person want that today is
833
00:34:31,780 --> 00:34:36,520
very difficult to do part of the problem
834
00:34:34,869 --> 00:34:39,119
is that the number of ways people
835
00:34:36,520 --> 00:34:41,409
gesture at you is very very large
836
00:34:39,119 --> 00:34:43,300
imagine all the hand gestures someone
837
00:34:41,409 --> 00:34:46,090
could conceivably use awesome
838
00:34:43,300 --> 00:34:47,950
slow down or go I'll stop the number of
839
00:34:46,090 --> 00:34:50,560
ways that people could gesture at you is
840
00:34:47,950 --> 00:34:53,110
just very very large and so it's
841
00:34:50,560 --> 00:34:55,660
difficult to collect enough data from
842
00:34:53,110 --> 00:34:58,120
enough thousands or tens of thousands or
843
00:34:55,660 --> 00:35:00,280
different people gesturing at you and
844
00:34:58,120 --> 00:35:02,830
all of these different ways to capture
845
00:35:00,280 --> 00:35:05,830
at the richness of human gestures so
846
00:35:02,830 --> 00:35:07,780
learning from a video to what this
847
00:35:05,830 --> 00:35:09,940
person one since I share somewhat
848
00:35:07,780 --> 00:35:11,860
complicated concept and that even people
849
00:35:09,940 --> 00:35:13,750
have a hard time figuring out sometimes
850
00:35:11,860 --> 00:35:16,300
what someone waving at your car wants
851
00:35:13,750 --> 00:35:18,430
and then second because this is a safety
852
00:35:16,300 --> 00:35:21,340
critical application you would want an
853
00:35:18,430 --> 00:35:23,080
AI that is extremely accurate in terms
854
00:35:21,340 --> 00:35:25,150
of figuring out there's a construction
855
00:35:23,080 --> 00:35:27,130
worker want you to stop or does he or
856
00:35:25,150 --> 00:35:30,850
she want you to go and that makes it
857
00:35:27,130 --> 00:35:34,060
harder for an AI system as well and so
858
00:35:30,850 --> 00:35:36,280
today if you collect just say ten
859
00:35:34,060 --> 00:35:37,690
thousand pictures of other cars many
860
00:35:36,280 --> 00:35:39,820
teams would be able to build an AI
861
00:35:37,690 --> 00:35:42,340
system that at least has a basic
862
00:35:39,820 --> 00:35:46,210
capability at detecting other cars
863
00:35:42,340 --> 00:35:48,190
in contrast even if you collect pictures
864
00:35:46,210 --> 00:35:50,080
or videos of ten thousand people is
865
00:35:48,190 --> 00:35:52,420
quite hard to track down ten thousand
866
00:35:50,080 --> 00:35:54,190
people waving at your car even with that
867
00:35:52,420 --> 00:35:57,130
data set I think it's quite hard today
868
00:35:54,190 --> 00:35:59,470
to build an AI system to recognize human
869
00:35:57,130 --> 00:36:01,450
intention from their gestures and at the
870
00:35:59,470 --> 00:36:03,580
very high level of accuracy needed in
871
00:36:01,450 --> 00:36:06,280
order to drive safely around these
872
00:36:03,580 --> 00:36:08,650
people so that's why today many sub
873
00:36:06,280 --> 00:36:10,240
driving car teams have some components
874
00:36:08,650 --> 00:36:12,400
for detecting other cars and they do
875
00:36:10,240 --> 00:36:15,070
rely on their technology to drive safely
876
00:36:12,400 --> 00:36:17,200
but very few self-driving car teams are
877
00:36:15,070 --> 00:36:19,990
trying to count on an AI system to
878
00:36:17,200 --> 00:36:22,300
recognize a huge diversity of human
879
00:36:19,990 --> 00:36:24,580
gestures and Counting just on that they
880
00:36:22,300 --> 00:36:26,920
drive safely around people let's look at
881
00:36:24,580 --> 00:36:29,830
one more example say you want to build
882
00:36:26,920 --> 00:36:32,470
an AI system to look at x-ray images and
883
00:36:29,830 --> 00:36:35,770
diagnose pneumonia so all of these are
884
00:36:32,470 --> 00:36:39,220
chest x-rays so the input a could be the
885
00:36:35,770 --> 00:36:41,050
x-ray image and the output B can be the
886
00:36:39,220 --> 00:36:43,480
diagnosis does this patient have
887
00:36:41,050 --> 00:36:46,060
pneumonia or not so that's something
888
00:36:43,480 --> 00:36:48,970
that a I can do something that a I
889
00:36:46,060 --> 00:36:52,120
cannot do would be to diagnose pneumonia
890
00:36:48,970 --> 00:36:54,400
from ten images of a medical textbook
891
00:36:52,120 --> 00:36:57,099
chapter explaining pneumonia
892
00:36:54,400 --> 00:36:59,799
a human can look at a small set of
893
00:36:57,099 --> 00:37:02,529
images maybe just a few dozen images and
894
00:36:59,799 --> 00:37:05,380
read a few paragraphs from a medical
895
00:37:02,529 --> 00:37:07,180
textbook and start to get a sense but I
896
00:37:05,380 --> 00:37:10,210
actually don't know given a medical
897
00:37:07,180 --> 00:37:12,549
textbook what is a and what is B or how
898
00:37:10,210 --> 00:37:13,869
to really pose this as an AI problem
899
00:37:12,549 --> 00:37:16,059
that I know how to write a piece of
900
00:37:13,869 --> 00:37:18,549
software to solve if all you have is
901
00:37:16,059 --> 00:37:21,010
just ten images and a few paragraphs of
902
00:37:18,549 --> 00:37:23,289
text that explain what pneumonia and a
903
00:37:21,010 --> 00:37:25,569
chest x-ray looks like whereas a young
904
00:37:23,289 --> 00:37:27,670
medical doctor might learn quite well
905
00:37:25,569 --> 00:37:29,710
reading a medical textbook and just
906
00:37:27,670 --> 00:37:33,309
looking at you know maybe dozens of
907
00:37:29,710 --> 00:37:34,869
images in contrast an AI system isn't
908
00:37:33,309 --> 00:37:36,910
really able to do that today
909
00:37:34,869 --> 00:37:38,680
so summarize here are some of the
910
00:37:36,910 --> 00:37:40,630
strengths and weaknesses of machine
911
00:37:38,680 --> 00:37:42,250
learning machine learning tends to work
912
00:37:40,630 --> 00:37:44,140
well when you're trying to learn a
913
00:37:42,250 --> 00:37:46,119
simple concept such as something that
914
00:37:44,140 --> 00:37:48,549
you could do with less than a second of
915
00:37:46,119 --> 00:37:51,490
mental thought and when there's lots of
916
00:37:48,549 --> 00:37:53,529
data available machine learning tends
917
00:37:51,490 --> 00:37:56,200
were poorly when you're trying to learn
918
00:37:53,529 --> 00:37:59,289
a complex concept from small amounts of
919
00:37:56,200 --> 00:38:01,960
data a second underappreciated weakness
920
00:37:59,289 --> 00:38:03,700
of AI is that it tends to do poorly when
921
00:38:01,960 --> 00:38:06,400
it's also perform on new types of data
922
00:38:03,700 --> 00:38:08,859
that's different than the data it has
923
00:38:06,400 --> 00:38:10,990
seen in your data set let me explain
924
00:38:08,859 --> 00:38:13,630
with an example say you built a
925
00:38:10,990 --> 00:38:15,990
supervised learning system that uses a
926
00:38:13,630 --> 00:38:19,029
to be to learn to diagnose pneumonia
927
00:38:15,990 --> 00:38:21,130
from images like these these are you
928
00:38:19,029 --> 00:38:23,980
know pretty high-quality chest x-ray
929
00:38:21,130 --> 00:38:26,410
images but now let's say you take this
930
00:38:23,980 --> 00:38:28,599
AI system and apply it at a different
931
00:38:26,410 --> 00:38:31,599
Hospital or a different Medical Center
932
00:38:28,599 --> 00:38:33,549
where maybe the x-ray technician somehow
933
00:38:31,599 --> 00:38:36,039
strangely had the patients always lied
934
00:38:33,549 --> 00:38:38,140
an angle or sometimes there are these
935
00:38:36,039 --> 00:38:40,650
defects not sure you can see the little
936
00:38:38,140 --> 00:38:43,779
scratches in the image these law other
937
00:38:40,650 --> 00:38:47,109
objects lying on top of the patient's if
938
00:38:43,779 --> 00:38:49,210
the AI system has learned from data like
939
00:38:47,109 --> 00:38:52,510
that on your left maybe taken from a
940
00:38:49,210 --> 00:38:54,970
high quality Medical Center and you take
941
00:38:52,510 --> 00:38:57,279
this AI system and apply it to a
942
00:38:54,970 --> 00:38:59,410
different Medical Center that generates
943
00:38:57,279 --> 00:39:01,690
images like those on the right then this
944
00:38:59,410 --> 00:39:04,109
performance would be quite poor as well
945
00:39:01,690 --> 00:39:06,579
a good AI team would be able to
946
00:39:04,109 --> 00:39:07,609
ameliorate or to reduce some of these
947
00:39:06,579 --> 00:39:10,880
problems
948
00:39:07,609 --> 00:39:12,589
but doing this is not that easy and this
949
00:39:10,880 --> 00:39:14,180
is one of the things that AI is actually
950
00:39:12,589 --> 00:39:17,599
much weaker than humans
951
00:39:14,180 --> 00:39:20,029
if a human has learned from images on
952
00:39:17,599 --> 00:39:22,430
the left they're much more likely to be
953
00:39:20,029 --> 00:39:23,809
able to adapt to images like those on
954
00:39:22,430 --> 00:39:25,940
the right as they figure out that the
955
00:39:23,809 --> 00:39:28,130
patient's just lying on an ankle but an
956
00:39:25,940 --> 00:39:30,890
AI system can be much less robust than
957
00:39:28,130 --> 00:39:32,930
human doctors in generalizing or
958
00:39:30,890 --> 00:39:35,269
freaking out what to do with new types
959
00:39:32,930 --> 00:39:37,819
of data like this I hope these examples
960
00:39:35,269 --> 00:39:40,910
are helping you hone your intuitions
961
00:39:37,819 --> 00:39:42,469
about what a I can and cannot do in case
962
00:39:40,910 --> 00:39:44,900
the boundary between what they can and
963
00:39:42,469 --> 00:39:46,430
cannot do so since fuzzy to you don't
964
00:39:44,900 --> 00:39:47,390
worry that's completely normal
965
00:39:46,430 --> 00:39:49,579
completely okay
966
00:39:47,390 --> 00:39:51,380
in fact even today I still can't open a
967
00:39:49,579 --> 00:39:53,509
project and immediately tell is
968
00:39:51,380 --> 00:39:56,359
something that's peaceful or not and I
969
00:39:53,509 --> 00:39:58,130
often still need reeks of small numbers
970
00:39:56,359 --> 00:40:00,349
of weeks of technical diligence before
971
00:39:58,130 --> 00:40:03,349
forming strong conviction about whether
972
00:40:00,349 --> 00:40:05,599
something is feasible or not but I hope
973
00:40:03,349 --> 00:40:07,729
that these examples can at least help
974
00:40:05,599 --> 00:40:09,950
you start imagining some things in your
975
00:40:07,729 --> 00:40:12,890
company that might be feasible and might
976
00:40:09,950 --> 00:40:15,259
be worth exploring more the next two
977
00:40:12,890 --> 00:40:17,539
videos after this are optional and are a
978
00:40:15,259 --> 00:40:18,890
non-technical description of whether
979
00:40:17,539 --> 00:40:20,869
neural networks and what is deep
980
00:40:18,890 --> 00:40:23,539
learning please feel free to watch those
981
00:40:20,869 --> 00:40:25,940
and then next week will go much more
982
00:40:23,539 --> 00:40:28,130
deeply into the process of what building
983
00:40:25,940 --> 00:40:31,269
and AI project would look like look
984
00:40:28,130 --> 00:40:31,269
forward to see you next week
985
00:40:34,580 --> 00:40:39,200
the terms deep learning and neural
986
00:40:37,460 --> 00:40:42,530
network are used almost interchangeably
987
00:40:39,200 --> 00:40:44,570
in AI and even though they're great for
988
00:40:42,530 --> 00:40:47,120
machine learning there's also been a bit
989
00:40:44,570 --> 00:40:49,820
of hype and bit of mystique about them
990
00:40:47,120 --> 00:40:52,520
this video will demystify deep learning
991
00:40:49,820 --> 00:40:54,830
so that you have a sense of what deep
992
00:40:52,520 --> 00:40:57,530
learning and neural networks really are
993
00:40:54,830 --> 00:40:59,990
let's use an example from demand
994
00:40:57,530 --> 00:41:03,320
prediction let's say you run a website
995
00:40:59,990 --> 00:41:05,900
that sells t-shirts and you want to know
996
00:41:03,320 --> 00:41:08,030
based on how you price the t-shirts
997
00:41:05,900 --> 00:41:09,980
how many units you expect to sell how
998
00:41:08,030 --> 00:41:12,170
many t-shirts you expect to sell you
999
00:41:09,980 --> 00:41:14,000
might then create a data set like this
1000
00:41:12,170 --> 00:41:16,910
where the higher the price of the
1001
00:41:14,000 --> 00:41:19,460
t-shirt the Lord to demand so you might
1002
00:41:16,910 --> 00:41:21,830
fit a straight line to this data showing
1003
00:41:19,460 --> 00:41:25,340
that as the price goes up the demand
1004
00:41:21,830 --> 00:41:28,250
goes down now demand can never go below
1005
00:41:25,340 --> 00:41:30,620
zero so maybe you say that the demand
1006
00:41:28,250 --> 00:41:32,390
will flatten out at zero and beyond a
1007
00:41:30,620 --> 00:41:35,960
certain point you expect you know pretty
1008
00:41:32,390 --> 00:41:39,260
much no one to buy any t-shirts it turns
1009
00:41:35,960 --> 00:41:42,230
out this blue line is maybe the simplest
1010
00:41:39,260 --> 00:41:46,660
possible neural network you have as
1011
00:41:42,230 --> 00:41:51,020
input the price ate and you wanted to
1012
00:41:46,660 --> 00:41:53,810
output the estimated demand B so the way
1013
00:41:51,020 --> 00:41:57,700
you would draw this as a new network is
1014
00:41:53,810 --> 00:42:00,890
that the price would be input to this
1015
00:41:57,700 --> 00:42:04,490
little round thing there and this little
1016
00:42:00,890 --> 00:42:07,460
round thing outputs yes me to demand in
1017
00:42:04,490 --> 00:42:10,850
these terminology of AI this little
1018
00:42:07,460 --> 00:42:12,860
round thing here is called a neuron or
1019
00:42:10,850 --> 00:42:16,400
sometimes it's called an artificial
1020
00:42:12,860 --> 00:42:19,340
neuron and oh it does is compute this
1021
00:42:16,400 --> 00:42:23,060
blue curve that I've drawn here on the
1022
00:42:19,340 --> 00:42:25,940
left this is maybe the simplest possible
1023
00:42:23,060 --> 00:42:28,010
neural network with a single artificial
1024
00:42:25,940 --> 00:42:31,040
neuron that just inputs the price and
1025
00:42:28,010 --> 00:42:33,770
outputs the estimated demand if you
1026
00:42:31,040 --> 00:42:36,980
think of this orange circle this
1027
00:42:33,770 --> 00:42:39,500
artificial neuron as a little Lego break
1028
00:42:36,980 --> 00:42:41,210
all the other neural network is is if
1029
00:42:39,500 --> 00:42:42,920
you take a lot of these Lego breaks and
1030
00:42:41,210 --> 00:42:45,950
stack them on top of each other until
1031
00:42:42,920 --> 00:42:48,140
you get a big how or a big network
1032
00:42:45,950 --> 00:42:51,170
of this Niraj let's look at a more
1033
00:42:48,140 --> 00:42:53,690
complex example suppose that instead of
1034
00:42:51,170 --> 00:42:57,020
knowing only the price of the t-shirts
1035
00:42:53,690 --> 00:42:59,270
you also have the shipping costs that
1036
00:42:57,020 --> 00:43:02,059
the customers will have to pay to get
1037
00:42:59,270 --> 00:43:05,319
the t-shirts maybe you spend more or
1038
00:43:02,059 --> 00:43:08,150
less on marketing in a given week and
1039
00:43:05,319 --> 00:43:11,390
you can also make the t-shirt out of a
1040
00:43:08,150 --> 00:43:14,450
thick heavy expensive cotton or a much
1041
00:43:11,390 --> 00:43:16,970
cheaper more lightweight material these
1042
00:43:14,450 --> 00:43:19,670
are some of the factors that you think
1043
00:43:16,970 --> 00:43:21,980
will affect the demand for your t-shirts
1044
00:43:19,670 --> 00:43:24,109
let's see what a more complex neural
1045
00:43:21,980 --> 00:43:25,599
network might look like you know that
1046
00:43:24,109 --> 00:43:30,230
your customers care a lot about
1047
00:43:25,599 --> 00:43:32,780
affordability so let's say you have one
1048
00:43:30,230 --> 00:43:36,319
neuron and let me draw this one in blue
1049
00:43:32,780 --> 00:43:39,020
whose job it is to estimate the
1050
00:43:36,319 --> 00:43:41,690
affordability of the t-shirts and
1051
00:43:39,020 --> 00:43:44,809
because affordability and so
1052
00:43:41,690 --> 00:43:46,549
affordability is mainly a function of
1053
00:43:44,809 --> 00:43:48,440
the price of the shirt and of the
1054
00:43:46,549 --> 00:43:50,630
shipping costs a second thing they'll
1055
00:43:48,440 --> 00:43:51,799
affect the demand for your teachers is
1056
00:43:50,630 --> 00:43:54,440
awareness
1057
00:43:51,799 --> 00:43:57,079
how much are consumers aware that you're
1058
00:43:54,440 --> 00:44:00,069
selling this t-shirt so the main thing
1059
00:43:57,079 --> 00:44:03,260
that affects awareness is going to be
1060
00:44:00,069 --> 00:44:06,230
your marketing so let me draw here a
1061
00:44:03,260 --> 00:44:08,390
second artificial neuron that inputs
1062
00:44:06,230 --> 00:44:10,220
your marketing budget how much you spent
1063
00:44:08,390 --> 00:44:14,770
on marketing and outputs
1064
00:44:10,220 --> 00:44:17,750
how aware are consumers of your t-shirt
1065
00:44:14,770 --> 00:44:20,109
finally the perceived quality of your
1066
00:44:17,750 --> 00:44:23,329
product will also affect demand and
1067
00:44:20,109 --> 00:44:25,730
perceived quality would be affected by
1068
00:44:23,329 --> 00:44:27,170
marketing if the marketing tries to
1069
00:44:25,730 --> 00:44:30,020
convince people this is a high quality
1070
00:44:27,170 --> 00:44:33,260
t-shirt and sometimes the price of
1071
00:44:30,020 --> 00:44:35,440
something also affects perceived quality
1072
00:44:33,260 --> 00:44:39,040
so I'm going to draw here a third
1073
00:44:35,440 --> 00:44:42,140
artificial neuron that inputs price
1074
00:44:39,040 --> 00:44:47,079
marketing and material and tries to
1075
00:44:42,140 --> 00:44:47,079
estimate the perceived quality
1076
00:44:47,599 --> 00:44:54,589
of your t-shirts finally now that the
1077
00:44:51,469 --> 00:44:57,619
earlier neurons these three blue neurons
1078
00:44:54,589 --> 00:44:58,999
have figured out how affordable how much
1079
00:44:57,619 --> 00:45:01,609
consumer awareness and what's a
1080
00:44:58,999 --> 00:45:03,949
perceived quality you can then have one
1081
00:45:01,609 --> 00:45:07,609
more neuron over here that takes us
1082
00:45:03,949 --> 00:45:12,170
input these three factors and outputs
1083
00:45:07,609 --> 00:45:15,769
the estimated demand so this is a neural
1084
00:45:12,170 --> 00:45:19,940
network and its job is to learn to map
1085
00:45:15,769 --> 00:45:25,279
from these four inputs that's the input
1086
00:45:19,940 --> 00:45:29,029
a to the output B to demand so it learns
1087
00:45:25,279 --> 00:45:31,579
this input output or a to be mapping
1088
00:45:29,029 --> 00:45:34,279
this is a fairly small neural network
1089
00:45:31,579 --> 00:45:36,769
with just four artificial neurons in
1090
00:45:34,279 --> 00:45:38,749
practice neural networks used today are
1091
00:45:36,769 --> 00:45:41,239
much larger more easily
1092
00:45:38,749 --> 00:45:44,630
thousands tens of thousands or even much
1093
00:45:41,239 --> 00:45:46,910
larger than that numbers of neurons now
1094
00:45:44,630 --> 00:45:48,619
there's just one final detail of this
1095
00:45:46,910 --> 00:45:51,440
description that I want to clean up
1096
00:45:48,619 --> 00:45:52,519
which is that in the way of describing
1097
00:45:51,440 --> 00:45:55,249
neural network
1098
00:45:52,519 --> 00:45:57,430
it was as if you had to figure out that
1099
00:45:55,249 --> 00:46:01,130
the key factors are affordability
1100
00:45:57,430 --> 00:46:03,349
awareness and perceived quality one of
1101
00:46:01,130 --> 00:46:05,959
the wonderful things about using neural
1102
00:46:03,349 --> 00:46:07,910
networks is that to train a neural
1103
00:46:05,959 --> 00:46:09,499
network in other words to build a
1104
00:46:07,910 --> 00:46:12,019
machine learning system using a neural
1105
00:46:09,499 --> 00:46:14,959
network all you have to do is give it
1106
00:46:12,019 --> 00:46:16,759
the input a and the upper B and it
1107
00:46:14,959 --> 00:46:19,999
figures out all of the things in the
1108
00:46:16,759 --> 00:46:22,910
middle by yourself so to build a neural
1109
00:46:19,999 --> 00:46:28,549
network what you would do is feed it
1110
00:46:22,910 --> 00:46:30,680
lots of data with the input a and have a
1111
00:46:28,549 --> 00:46:33,229
neural network that just looks like this
1112
00:46:30,680 --> 00:46:35,749
with a few blue neurons feeding to a
1113
00:46:33,229 --> 00:46:38,569
yellow output neuron and then you have
1114
00:46:35,749 --> 00:46:41,869
two given data with the demand B as well
1115
00:46:38,569 --> 00:46:45,170
and it's the software's job to figure
1116
00:46:41,869 --> 00:46:47,839
out what these blue neurons should be
1117
00:46:45,170 --> 00:46:49,880
computing so that it can completely
1118
00:46:47,839 --> 00:46:52,699
automatically learn the most accurate
1119
00:46:49,880 --> 00:46:55,729
possible function mapping from the input
1120
00:46:52,699 --> 00:46:58,219
to the output B and it turns out that if
1121
00:46:55,729 --> 00:47:00,580
you give this enough data and train a
1122
00:46:58,219 --> 00:47:02,920
neural network that is big enough there
1123
00:47:00,580 --> 00:47:06,370
can do an incredibly good job mapping
1124
00:47:02,920 --> 00:47:09,940
from inputs a to uppers beam so that's a
1125
00:47:06,370 --> 00:47:11,950
neural network is a group of artificial
1126
00:47:09,940 --> 00:47:14,080
neurons each of which computes a
1127
00:47:11,950 --> 00:47:15,880
relatively simple function but when you
1128
00:47:14,080 --> 00:47:18,070
stack enough of them together like Lego
1129
00:47:15,880 --> 00:47:20,050
breaks they can compute incredibly
1130
00:47:18,070 --> 00:47:22,660
complicated functions that give you very
1131
00:47:20,050 --> 00:47:25,600
accurate mappings from the input a to
1132
00:47:22,660 --> 00:47:27,970
the output B now in this video you saw
1133
00:47:25,600 --> 00:47:30,520
an example of neural networks apply to
1134
00:47:27,970 --> 00:47:32,830
demand prediction let's go onto the next
1135
00:47:30,520 --> 00:47:37,290
video to see a more complex example of
1136
00:47:32,830 --> 00:47:37,290
new networks apply to face recognition
1137
00:47:40,380 --> 00:47:45,570
in the last video you saw how a neural
1138
00:47:43,380 --> 00:47:48,120
network can be applied to demand
1139
00:47:45,570 --> 00:47:49,620
prediction but how can a new network
1140
00:47:48,120 --> 00:47:51,210
look at the picture and figure out
1141
00:47:49,620 --> 00:47:53,580
what's in the picture or listen to an
1142
00:47:51,210 --> 00:47:55,800
audio clip and understand what is said
1143
00:47:53,580 --> 00:47:57,690
in an audio clip let's take a look at a
1144
00:47:55,800 --> 00:48:00,780
more complex example of applying a
1145
00:47:57,690 --> 00:48:02,460
neural network to face recognition say
1146
00:48:00,780 --> 00:48:05,760
you want to build a system to recognize
1147
00:48:02,460 --> 00:48:08,430
people from pictures how can a piece of
1148
00:48:05,760 --> 00:48:10,530
software look at this picture and figure
1149
00:48:08,430 --> 00:48:13,110
out the identity of the person in it
1150
00:48:10,530 --> 00:48:15,240
let's zoom into a little square like
1151
00:48:13,110 --> 00:48:17,910
that to better understand how a computer
1152
00:48:15,240 --> 00:48:20,820
sees pictures where you and I see a
1153
00:48:17,910 --> 00:48:24,180
human eye a computer is that sees that
1154
00:48:20,820 --> 00:48:26,460
it sees this grid of pixel brightness
1155
00:48:24,180 --> 00:48:29,340
values that tells it for each of the
1156
00:48:26,460 --> 00:48:32,280
pixels in the image how bright is that
1157
00:48:29,340 --> 00:48:35,670
pixel if it were a black and white or a
1158
00:48:32,280 --> 00:48:38,250
grayscale image then each pixel would
1159
00:48:35,670 --> 00:48:40,590
correspond to a single number telling
1160
00:48:38,250 --> 00:48:43,050
you how bright is that pixel if is a
1161
00:48:40,590 --> 00:48:44,690
color image then each pixel would
1162
00:48:43,050 --> 00:48:47,370
actually have three numbers
1163
00:48:44,690 --> 00:48:51,150
corresponding to how bright are the red
1164
00:48:47,370 --> 00:48:54,420
green and blue elements of that pixel so
1165
00:48:51,150 --> 00:48:57,240
the new networks job is to take us input
1166
00:48:54,420 --> 00:48:59,790
a lot of numbers like these and tell you
1167
00:48:57,240 --> 00:49:02,340
the name of the person in the picture in
1168
00:48:59,790 --> 00:49:04,580
the last video you saw how a new network
1169
00:49:02,340 --> 00:49:07,110
can take as input four numbers
1170
00:49:04,580 --> 00:49:09,870
corresponding to the price shipping cost
1171
00:49:07,110 --> 00:49:13,650
amount of marketing and cloth material
1172
00:49:09,870 --> 00:49:16,080
of a t-shirt and output demand in this
1173
00:49:13,650 --> 00:49:18,900
example the neural network just has to
1174
00:49:16,080 --> 00:49:21,720
input a lot more numbers corresponding
1175
00:49:18,900 --> 00:49:25,170
to all of the pixel brightness values of
1176
00:49:21,720 --> 00:49:29,130
this picture if the resolution of this
1177
00:49:25,170 --> 00:49:32,190
picture is 1000 pixels by 1000 pixels
1178
00:49:29,130 --> 00:49:34,530
then that's a million pixels so if it
1179
00:49:32,190 --> 00:49:36,660
were a black and white or grayscale
1180
00:49:34,530 --> 00:49:40,470
image this neural network was take as
1181
00:49:36,660 --> 00:49:43,260
input a million numbers corresponding to
1182
00:49:40,470 --> 00:49:47,460
the brightness of all 1 million pixels
1183
00:49:43,260 --> 00:49:50,660
in this image or it was a color image it
1184
00:49:47,460 --> 00:49:53,860
would take as input 3 million numbers
1185
00:49:50,660 --> 00:49:56,440
corresponding to the red green and blue
1186
00:49:53,860 --> 00:49:59,650
values of each of these 1 million pixels
1187
00:49:56,440 --> 00:50:02,020
in this image similar to before you will
1188
00:49:59,650 --> 00:50:05,470
have many many of these artificial
1189
00:50:02,020 --> 00:50:07,810
neurons computing various values and
1190
00:50:05,470 --> 00:50:09,940
it's not your job to figure out what
1191
00:50:07,810 --> 00:50:13,150
these neurons should compute the new
1192
00:50:09,940 --> 00:50:16,000
network will figure it out by itself but
1193
00:50:13,150 --> 00:50:18,460
typically when you give it an image the
1194
00:50:16,000 --> 00:50:20,650
neurons in the earlier parts of the
1195
00:50:18,460 --> 00:50:23,740
neural network will learn to detect
1196
00:50:20,650 --> 00:50:26,020
edges in pictures and then little bit
1197
00:50:23,740 --> 00:50:28,420
later or learn to detect parts of
1198
00:50:26,020 --> 00:50:30,610
objects so may learn to detect eyes and
1199
00:50:28,420 --> 00:50:32,880
noses and the shape of cheeks in the
1200
00:50:30,610 --> 00:50:35,470
shape of models and then the later
1201
00:50:32,880 --> 00:50:37,600
neurons further to the right will learn
1202
00:50:35,470 --> 00:50:40,450
to detect different shapes of faces and
1203
00:50:37,600 --> 00:50:43,420
it will finally put all this together to
1204
00:50:40,450 --> 00:50:46,060
output the identity of the person indeed
1205
00:50:43,420 --> 00:50:48,160
image and again part of the magic of
1206
00:50:46,060 --> 00:50:50,680
neural networks is that you don't really
1207
00:50:48,160 --> 00:50:53,290
need to worry about what it is doing in
1208
00:50:50,680 --> 00:50:56,380
the middle all you need to do is given a
1209
00:50:53,290 --> 00:50:59,890
lot of data of pictures like this a as
1210
00:50:56,380 --> 00:51:02,350
well as of the correct identity B and
1211
00:50:59,890 --> 00:51:03,640
the learning algorithm will figure out
1212
00:51:02,350 --> 00:51:05,710
by itself
1213
00:51:03,640 --> 00:51:08,740
what each of these neurons in the middle
1214
00:51:05,710 --> 00:51:10,900
should be computing congratulations on
1215
00:51:08,740 --> 00:51:13,690
finishing all the videos for this week
1216
00:51:10,900 --> 00:51:16,690
you now know how machine learning and
1217
00:51:13,690 --> 00:51:18,910
data signs work I look forward to seeing
1218
00:51:16,690 --> 00:51:21,100
you in next week's videos as well where
1219
00:51:18,910 --> 00:51:23,890
you learn how to build your own machine
1220
00:51:21,100 --> 00:51:26,250
learning or data science project see you
1221
00:51:23,890 --> 00:51:26,250
next week
1222
00:51:28,420 --> 00:51:34,559
machine learning algorithms can learn
1223
00:51:30,670 --> 00:51:37,299
input the output or a to be mappings so
1224
00:51:34,559 --> 00:51:39,670
how do you build a machine learning
1225
00:51:37,299 --> 00:51:40,930
project in this video you learn what is
1226
00:51:39,670 --> 00:51:43,779
the workflow of machine learning
1227
00:51:40,930 --> 00:51:45,700
projects let's take a look as a running
1228
00:51:43,779 --> 00:51:47,829
example I'm going to use speech
1229
00:51:45,700 --> 00:51:50,140
recognition so some of you may have an
1230
00:51:47,829 --> 00:51:53,440
Amazon echo or a Google home or Apple
1231
00:51:50,140 --> 00:51:55,749
sory device or a Baidu to iOS device in
1232
00:51:53,440 --> 00:51:57,579
your homes some years back I done some
1233
00:51:55,749 --> 00:52:00,009
work on Google's speech recognition
1234
00:51:57,579 --> 00:52:02,619
system and it also led by juice doer oh
1235
00:52:00,009 --> 00:52:05,769
s project and today I actually have a
1236
00:52:02,619 --> 00:52:08,470
Amazon echo in my kitchen so every time
1237
00:52:05,769 --> 00:52:10,660
I'm balding an egg I will say Alexa set
1238
00:52:08,470 --> 00:52:12,730
timer for three minutes and then it lets
1239
00:52:10,660 --> 00:52:15,460
you know when the Freedmen's are up and
1240
00:52:12,730 --> 00:52:17,499
my eggs are ready so how do you build a
1241
00:52:15,460 --> 00:52:20,349
speech recognition system that can
1242
00:52:17,499 --> 00:52:23,859
recognize when you say Alexa or hey
1243
00:52:20,349 --> 00:52:25,660
Google or hey Siri or hello Baidu let's
1244
00:52:23,859 --> 00:52:28,269
go through the key steps of a machine
1245
00:52:25,660 --> 00:52:30,069
learning project and just for simplicity
1246
00:52:28,269 --> 00:52:33,249
I'm going to use Amazon echo or
1247
00:52:30,069 --> 00:52:35,559
detecting the Alexa keyword as this
1248
00:52:33,249 --> 00:52:36,999
running example if you want to build an
1249
00:52:35,559 --> 00:52:39,190
AI system or build a machine learning
1250
00:52:36,999 --> 00:52:42,009
system to figure out when a user has
1251
00:52:39,190 --> 00:52:44,710
said the word Alexa the first step is to
1252
00:52:42,009 --> 00:52:47,890
collect data so that means you go around
1253
00:52:44,710 --> 00:52:50,619
and get some people to say the words
1254
00:52:47,890 --> 00:52:52,869
Alexa for you and you record the audio
1255
00:52:50,619 --> 00:52:55,660
of that and you'll also get a bunch of
1256
00:52:52,869 --> 00:52:57,819
people to say other words like hello or
1257
00:52:55,660 --> 00:53:01,299
say lots of other words and record the
1258
00:52:57,819 --> 00:53:03,730
audio of that as well having collected a
1259
00:53:01,299 --> 00:53:06,009
lot of audio data a lot of these audio
1260
00:53:03,730 --> 00:53:08,589
clips that people saying either Alexa or
1261
00:53:06,009 --> 00:53:11,230
saying other things step two is to then
1262
00:53:08,589 --> 00:53:13,239
train the model and this means you will
1263
00:53:11,230 --> 00:53:16,900
use a machine learning algorithm to
1264
00:53:13,239 --> 00:53:19,690
learn an input to output or a to be
1265
00:53:16,900 --> 00:53:22,029
mapping where the input a would be an
1266
00:53:19,690 --> 00:53:24,579
audio clip and in the case of the first
1267
00:53:22,029 --> 00:53:27,789
audio clip above hopefully it will tell
1268
00:53:24,579 --> 00:53:30,489
you that the user said Alexa and in the
1269
00:53:27,789 --> 00:53:33,339
case of audio clip two shown on the
1270
00:53:30,489 --> 00:53:35,829
right hopefully the system will learn to
1271
00:53:33,339 --> 00:53:39,279
recognize that the user has said hello
1272
00:53:35,829 --> 00:53:41,530
whenever an AI team starts to train the
1273
00:53:39,279 --> 00:53:43,600
model meaning to learn the eight
1274
00:53:41,530 --> 00:53:45,130
your input output mapping what happens
1275
00:53:43,600 --> 00:53:47,220
pretty much every time is the first
1276
00:53:45,130 --> 00:53:49,870
attempt doesn't work well and so
1277
00:53:47,220 --> 00:53:52,780
invariably the team will need to try
1278
00:53:49,870 --> 00:53:54,220
many times or in a I recall this iterate
1279
00:53:52,780 --> 00:53:56,590
many times you have to iterate many
1280
00:53:54,220 --> 00:53:59,920
times until hopefully the model looks
1281
00:53:56,590 --> 00:54:02,470
like is good enough the third step is to
1282
00:53:59,920 --> 00:54:04,360
then actually deploy the model and what
1283
00:54:02,470 --> 00:54:06,850
that means is you put this AI software
1284
00:54:04,360 --> 00:54:09,070
into an actual small speaker and ship it
1285
00:54:06,850 --> 00:54:11,860
to either a small group of test users or
1286
00:54:09,070 --> 00:54:14,470
to a large group of users what happens
1287
00:54:11,860 --> 00:54:16,510
in a lot of AI products is that when you
1288
00:54:14,470 --> 00:54:19,000
ship it you see that it starts getting
1289
00:54:16,510 --> 00:54:22,060
new data and it may not work as well as
1290
00:54:19,000 --> 00:54:24,100
you had initially hoped so for example I
1291
00:54:22,060 --> 00:54:26,020
am from the UK so I'm going to pick on
1292
00:54:24,100 --> 00:54:28,150
the British but let's say you had
1293
00:54:26,020 --> 00:54:31,330
trained your speech recognition system
1294
00:54:28,150 --> 00:54:33,910
on American accented speakers and you
1295
00:54:31,330 --> 00:54:36,250
then ship this small speaker to the UK
1296
00:54:33,910 --> 00:54:38,740
and you start having British accent -
1297
00:54:36,250 --> 00:54:40,930
people say Alexa then you may find that
1298
00:54:38,740 --> 00:54:43,210
it doesn't recognize their speech as
1299
00:54:40,930 --> 00:54:46,150
well as you had hoped and when that
1300
00:54:43,210 --> 00:54:49,450
happens hopefully you can get data back
1301
00:54:46,150 --> 00:54:51,550
of cases such as maybe British accented
1302
00:54:49,450 --> 00:54:53,560
speakers was not working as well as
1303
00:54:51,550 --> 00:54:56,440
you're hoping and then use this data to
1304
00:54:53,560 --> 00:54:58,840
maintain and to update the model so to
1305
00:54:56,440 --> 00:55:02,860
summarize the key steps of a machine
1306
00:54:58,840 --> 00:55:05,170
learning project are to collect data to
1307
00:55:02,860 --> 00:55:08,590
train the model that a to be mapping and
1308
00:55:05,170 --> 00:55:10,810
then to deploy the model and throughout
1309
00:55:08,590 --> 00:55:12,940
these steps there is often a lot of
1310
00:55:10,810 --> 00:55:14,590
iteration meaning fine tuning or
1311
00:55:12,940 --> 00:55:16,570
adapting the model to work better or
1312
00:55:14,590 --> 00:55:19,180
getting data back even after you've
1313
00:55:16,570 --> 00:55:21,280
shifted to hopefully make the product
1314
00:55:19,180 --> 00:55:23,110
better which may or may not be possible
1315
00:55:21,280 --> 00:55:24,910
depending on whether you're able to get
1316
00:55:23,110 --> 00:55:27,130
data back let's look at these three
1317
00:55:24,910 --> 00:55:29,200
steps and see how they apply on the
1318
00:55:27,130 --> 00:55:31,840
different projects on building a key
1319
00:55:29,200 --> 00:55:33,550
component of a self-driving car so
1320
00:55:31,840 --> 00:55:35,740
remember the key steps that collect data
1321
00:55:33,550 --> 00:55:37,990
at Raymond and deploy model since we
1322
00:55:35,740 --> 00:55:39,490
revisit these three steps on the next
1323
00:55:37,990 --> 00:55:41,410
slide let's say you're building a
1324
00:55:39,490 --> 00:55:43,480
self-driving car one of the key
1325
00:55:41,410 --> 00:55:45,880
components is a self-driving car is a
1326
00:55:43,480 --> 00:55:48,070
machine learning algorithm that takes as
1327
00:55:45,880 --> 00:55:49,990
input say a picture of what's in front
1328
00:55:48,070 --> 00:55:52,600
of your car and tells you where are the
1329
00:55:49,990 --> 00:55:54,080
other calls so what's the first step of
1330
00:55:52,600 --> 00:55:56,150
building this machine
1331
00:55:54,080 --> 00:55:58,550
learning system hopefully you remember
1332
00:55:56,150 --> 00:56:03,410
from the last night that the first step
1333
00:55:58,550 --> 00:56:04,940
was to collect data so if you go is have
1334
00:56:03,410 --> 00:56:07,850
a machine learning algorithm they could
1335
00:56:04,940 --> 00:56:09,950
take us input an image and output the
1336
00:56:07,850 --> 00:56:12,710
position of other cause the data you
1337
00:56:09,950 --> 00:56:15,200
need to collect would be both images as
1338
00:56:12,710 --> 00:56:18,050
well as position of other costs that you
1339
00:56:15,200 --> 00:56:19,940
want to a our system to output so let's
1340
00:56:18,050 --> 00:56:23,600
say you start off with a few pictures
1341
00:56:19,940 --> 00:56:25,940
like this these are the inputs a to the
1342
00:56:23,600 --> 00:56:28,370
machine learning algorithm you need to
1343
00:56:25,940 --> 00:56:30,050
also tell it what is the output B you
1344
00:56:28,370 --> 00:56:33,080
would want and so for each of these
1345
00:56:30,050 --> 00:56:36,020
pictures you would draw a rectangle
1346
00:56:33,080 --> 00:56:39,650
around the cause in the picture that you
1347
00:56:36,020 --> 00:56:41,780
wanted to detect and on this line I'm
1348
00:56:39,650 --> 00:56:44,570
hand drawing these rectangles but in
1349
00:56:41,780 --> 00:56:46,880
practice you will use some software that
1350
00:56:44,570 --> 00:56:48,920
lets you draw perfect rectangles rather
1351
00:56:46,880 --> 00:56:51,590
than these hand-drawn ones and then
1352
00:56:48,920 --> 00:56:54,050
having created this data set what was
1353
00:56:51,590 --> 00:56:56,930
the second step hope you remember that
1354
00:56:54,050 --> 00:57:01,070
the second step was to train them
1355
00:56:56,930 --> 00:57:03,620
although now invariably when you're AI
1356
00:57:01,070 --> 00:57:05,360
engineers start training a model they'll
1357
00:57:03,620 --> 00:57:07,490
find initially that it doesn't work that
1358
00:57:05,360 --> 00:57:09,710
well for example given this picture
1359
00:57:07,490 --> 00:57:12,230
maybe the software the first few tries
1360
00:57:09,710 --> 00:57:14,420
things that that is a car and it's only
1361
00:57:12,230 --> 00:57:16,460
by iterating many times that you
1362
00:57:14,420 --> 00:57:18,530
hopefully get a better result like
1363
00:57:16,460 --> 00:57:21,410
figuring out that that is where the car
1364
00:57:18,530 --> 00:57:25,010
actually is finally what was the third
1365
00:57:21,410 --> 00:57:27,170
step it was to deploy the model of
1366
00:57:25,010 --> 00:57:29,600
course in the self-driving world is
1367
00:57:27,170 --> 00:57:31,910
important to treat safety as number one
1368
00:57:29,600 --> 00:57:34,220
and deploy model or to test the model
1369
00:57:31,910 --> 00:57:36,860
only in ways they can preserve safety
1370
00:57:34,220 --> 00:57:39,110
but when you put the software in cars on
1371
00:57:36,860 --> 00:57:41,750
the road you may find that there are new
1372
00:57:39,110 --> 00:57:43,730
types of vehicles say golf cause that
1373
00:57:41,750 --> 00:57:46,520
the software isn't detecting very well
1374
00:57:43,730 --> 00:57:48,860
and so you get data back say pictures of
1375
00:57:46,520 --> 00:57:50,840
these golf cars use the new data to
1376
00:57:48,860 --> 00:57:53,410
maintain and update the model so that
1377
00:57:50,840 --> 00:57:56,120
hopefully you can have your AI software
1378
00:57:53,410 --> 00:57:58,580
continually get better and better to the
1379
00:57:56,120 --> 00:58:00,860
point where you end up with a software
1380
00:57:58,580 --> 00:58:03,890
that can do a pretty good job detecting
1381
00:58:00,860 --> 00:58:05,630
other costs from pictures like these in
1382
00:58:03,890 --> 00:58:06,850
this video you learn what are the key
1383
00:58:05,630 --> 00:58:09,280
steps of a machine
1384
00:58:06,850 --> 00:58:11,200
project which ought to collect data to
1385
00:58:09,280 --> 00:58:13,900
train them although and then to deploy
1386
00:58:11,200 --> 00:58:15,400
the model NYX let's take a look at
1387
00:58:13,900 --> 00:58:17,290
whether the key steps or what does it
1388
00:58:15,400 --> 00:58:20,610
work though of a data science project
1389
00:58:17,290 --> 00:58:20,610
let's go onto the next video
1390
00:58:23,040 --> 00:58:27,570
unlike a machine learning project the
1391
00:58:25,620 --> 00:58:30,390
output of a data science project is
1392
00:58:27,570 --> 00:58:32,100
often a set of actionable insights of
1393
00:58:30,390 --> 00:58:34,710
insights that may cause you to do things
1394
00:58:32,100 --> 00:58:36,240
differently so data science projects
1395
00:58:34,710 --> 00:58:38,070
have a different workflow the machine
1396
00:58:36,240 --> 00:58:39,690
learning projects let's take a look at
1397
00:58:38,070 --> 00:58:42,840
one of the steps of a data science
1398
00:58:39,690 --> 00:58:45,510
project as our running example let's say
1399
00:58:42,840 --> 00:58:48,060
you want to optimize a sales funnel say
1400
00:58:45,510 --> 00:58:50,760
you run a ecommerce or online shopping
1401
00:58:48,060 --> 00:58:52,740
website that sells coffee mugs and so
1402
00:58:50,760 --> 00:58:54,420
for a user to buy a coffee mug from you
1403
00:58:52,740 --> 00:58:56,550
there's a sequence of steps they'll
1404
00:58:54,420 --> 00:58:59,400
usually follow first they'll visit your
1405
00:58:56,550 --> 00:59:01,770
website and take a look at the different
1406
00:58:59,400 --> 00:59:04,380
coffee mugs on offer then eventually if
1407
00:59:01,770 --> 00:59:06,300
you get to a product page and then
1408
00:59:04,380 --> 00:59:08,010
they'll have to put it into the shopping
1409
00:59:06,300 --> 00:59:10,080
cart and go to the shopping cart page
1410
00:59:08,010 --> 00:59:13,740
and then they'll finally have to
1411
00:59:10,080 --> 00:59:15,990
checkout so if you want to optimize the
1412
00:59:13,740 --> 00:59:17,940
sales funnel to make sure that as many
1413
00:59:15,990 --> 00:59:20,400
people as possible get through all of
1414
00:59:17,940 --> 00:59:22,800
these steps how can you use data signs
1415
00:59:20,400 --> 00:59:24,900
to help with this problem let's look at
1416
00:59:22,800 --> 00:59:28,500
the key steps of a data science project
1417
00:59:24,900 --> 00:59:30,360
the first step is to collect data so on
1418
00:59:28,500 --> 00:59:32,940
a website like the one we saw you may
1419
00:59:30,360 --> 00:59:35,640
have a data set that's forced when
1420
00:59:32,940 --> 00:59:38,340
different users go to different web
1421
00:59:35,640 --> 00:59:40,710
pages in this simple example I'm
1422
00:59:38,340 --> 00:59:42,810
assuming that you can figure out the
1423
00:59:40,710 --> 00:59:44,820
country that the users are coming from
1424
00:59:42,810 --> 00:59:47,220
for example by looking at their
1425
00:59:44,820 --> 00:59:49,350
computers address called an IP address
1426
00:59:47,220 --> 00:59:52,410
and figuring out what is the country
1427
00:59:49,350 --> 00:59:54,540
from which they're originating but in
1428
00:59:52,410 --> 00:59:56,760
practice you can usually get quite a bit
1429
00:59:54,540 --> 00:59:59,190
more data about users than just what
1430
00:59:56,760 --> 01:00:01,560
country they're from the second step is
1431
00:59:59,190 --> 01:00:03,330
to then analyze the data your data
1432
01:00:01,560 --> 01:00:04,800
science team may have a lot of ideas
1433
01:00:03,330 --> 01:00:07,890
about what is affecting the performance
1434
01:00:04,800 --> 01:00:10,080
of your sales funnel for example they
1435
01:00:07,890 --> 01:00:12,090
may think that overseas customers are
1436
01:00:10,080 --> 01:00:14,400
scared off by the International shipping
1437
01:00:12,090 --> 01:00:15,930
costs which is why a lot of people go to
1438
01:00:14,400 --> 01:00:18,750
the checkout page but don't actually
1439
01:00:15,930 --> 01:00:21,120
check out and if that's true then you
1440
01:00:18,750 --> 01:00:22,740
might think about whether to put part of
1441
01:00:21,120 --> 01:00:25,230
shipping costs into the actual product
1442
01:00:22,740 --> 01:00:27,390
costs or your data science team may
1443
01:00:25,230 --> 01:00:29,340
think there are blips in the data
1444
01:00:27,390 --> 01:00:30,900
whenever there's a holiday maybe more
1445
01:00:29,340 --> 01:00:32,910
people will shop around the holidays
1446
01:00:30,900 --> 01:00:34,590
because the bank gives or maybe fewer
1447
01:00:32,910 --> 01:00:36,829
people will shop around the holidays
1448
01:00:34,590 --> 01:00:38,450
because they're staying home rather than
1449
01:00:36,829 --> 01:00:40,789
you know sometimes shopping from their
1450
01:00:38,450 --> 01:00:43,640
work computers and in some countries
1451
01:00:40,789 --> 01:00:45,979
there may be time of day blips where in
1452
01:00:43,640 --> 01:00:47,869
countries that observe a siesta so a
1453
01:00:45,979 --> 01:00:50,180
time of rest like an afternoon rest
1454
01:00:47,869 --> 01:00:51,920
there may be fewer shoppers online and
1455
01:00:50,180 --> 01:00:53,660
so your sales may go down and they might
1456
01:00:51,920 --> 01:00:56,059
didn't suggest that you should spend
1457
01:00:53,660 --> 01:00:57,920
fewer advertising dollars during the
1458
01:00:56,059 --> 01:01:01,519
period of CES there because fewer people
1459
01:00:57,920 --> 01:01:03,200
will go online to buy at that time so a
1460
01:01:01,519 --> 01:01:05,930
good data science team may have many
1461
01:01:03,200 --> 01:01:08,269
ideas and so they try many ideas or will
1462
01:01:05,930 --> 01:01:10,609
say anyway many times to get good
1463
01:01:08,269 --> 01:01:12,979
insights finally the data science team
1464
01:01:10,609 --> 01:01:16,729
will destroy these insights down to a
1465
01:01:12,979 --> 01:01:18,109
smaller number of hypotheses about ideas
1466
01:01:16,729 --> 01:01:20,029
of what could be going wrong what could
1467
01:01:18,109 --> 01:01:22,749
be going poorly as well as a smaller
1468
01:01:20,029 --> 01:01:24,680
number of suggested actions such as
1469
01:01:22,749 --> 01:01:26,539
incorporating shipping costs into the
1470
01:01:24,680 --> 01:01:29,180
product cost rather than having it as a
1471
01:01:26,539 --> 01:01:31,880
separate line item when you take some of
1472
01:01:29,180 --> 01:01:33,619
these suggested actions and deploy these
1473
01:01:31,880 --> 01:01:36,289
changes to your website you then start
1474
01:01:33,619 --> 01:01:37,849
to get new data back as users behave
1475
01:01:36,289 --> 01:01:39,920
differently now that you advertise
1476
01:01:37,849 --> 01:01:41,989
differently at the time of Siesta or of
1477
01:01:39,920 --> 01:01:43,640
a different checkout policy and then
1478
01:01:41,989 --> 01:01:46,130
your data science team can continue to
1479
01:01:43,640 --> 01:01:48,349
collect data and reanalyze the new data
1480
01:01:46,130 --> 01:01:50,509
periodically so see if they can come up
1481
01:01:48,349 --> 01:01:52,910
with even better hypotheses or even
1482
01:01:50,509 --> 01:01:54,650
better actions over time so the key
1483
01:01:52,910 --> 01:01:58,279
steps of a data science project are to
1484
01:01:54,650 --> 01:02:01,130
collect the data to analyze the data and
1485
01:01:58,279 --> 01:02:03,769
then to suggest hypotheses and actions
1486
01:02:01,130 --> 01:02:05,359
and then to continue to get the data
1487
01:02:03,769 --> 01:02:08,329
back and we analyze the data
1488
01:02:05,359 --> 01:02:10,549
periodically let's take this framework
1489
01:02:08,329 --> 01:02:14,420
and apply it to a new problem to
1490
01:02:10,549 --> 01:02:16,640
optimizing a manufacturing line and so
1491
01:02:14,420 --> 01:02:19,219
we'll take these three steps and use
1492
01:02:16,640 --> 01:02:20,930
them on the next slide as well let's say
1493
01:02:19,219 --> 01:02:22,789
you run the factories as manufacturing
1494
01:02:20,930 --> 01:02:24,829
thousands of coffee mugs a month for
1495
01:02:22,789 --> 01:02:27,709
sale and you want to optimize the
1496
01:02:24,829 --> 01:02:29,779
manufacturing line so these are the key
1497
01:02:27,709 --> 01:02:32,329
steps in manufacturing coffee maps step
1498
01:02:29,779 --> 01:02:34,420
one is the mixer clay so make sure the
1499
01:02:32,329 --> 01:02:37,009
appropriate amount of water is added
1500
01:02:34,420 --> 01:02:40,279
step two is to take the screen and to
1501
01:02:37,009 --> 01:02:42,650
shape the mugs then you have to add the
1502
01:02:40,279 --> 01:02:45,859
glaze so that the coloring protective
1503
01:02:42,650 --> 01:02:48,890
cover then you have to heat this mug and
1504
01:02:45,859 --> 01:02:50,349
we call that firing the kiln and finally
1505
01:02:48,890 --> 01:02:51,910
you would inspect the
1506
01:02:50,349 --> 01:02:53,799
to make sure there aren't dents in the
1507
01:02:51,910 --> 01:02:56,589
mug and it isn't cracked before you
1508
01:02:53,799 --> 01:02:59,019
should put a customer's so a common
1509
01:02:56,589 --> 01:03:01,179
problem in manufacturing is to optimize
1510
01:02:59,019 --> 01:03:03,609
the yield of this manufacturing line to
1511
01:03:01,179 --> 01:03:05,919
make sure that as you damage coffee mugs
1512
01:03:03,609 --> 01:03:07,359
get produced as possible because those
1513
01:03:05,919 --> 01:03:09,880
are coffee rust you have to throw away
1514
01:03:07,359 --> 01:03:11,979
resulting in time and material waste
1515
01:03:09,880 --> 01:03:13,839
what's the first step of a data science
1516
01:03:11,979 --> 01:03:15,579
project I hope you remember from the
1517
01:03:13,839 --> 01:03:19,089
last slide - the first step is to
1518
01:03:15,579 --> 01:03:21,609
collect data so for example you may save
1519
01:03:19,089 --> 01:03:24,429
data about the different batches of clay
1520
01:03:21,609 --> 01:03:26,140
that you've mix such as who supply the
1521
01:03:24,429 --> 01:03:27,939
clay and how long did you mix it or
1522
01:03:26,140 --> 01:03:29,499
maybe how much moisture wasn't it clear
1523
01:03:27,939 --> 01:03:31,539
how much water did you add you might
1524
01:03:29,499 --> 01:03:33,910
also collect data about the different
1525
01:03:31,539 --> 01:03:36,699
batches of mugs you made so how much
1526
01:03:33,910 --> 01:03:38,890
humidity wasn't that badge what was the
1527
01:03:36,699 --> 01:03:41,650
temperature in the kiln and how long did
1528
01:03:38,890 --> 01:03:44,109
you fire it in the kiln given all this
1529
01:03:41,650 --> 01:03:46,719
data you would then ask the data science
1530
01:03:44,109 --> 01:03:49,269
team to analyze the data and they would
1531
01:03:46,719 --> 01:03:51,519
ask before it read many times to get
1532
01:03:49,269 --> 01:03:54,069
good insights and so they might find out
1533
01:03:51,519 --> 01:03:56,049
for example that whenever the humidity
1534
01:03:54,069 --> 01:03:58,029
is too low and the kiln temperature is
1535
01:03:56,049 --> 01:03:59,919
too hot that it cracks in the mug or
1536
01:03:58,029 --> 01:04:02,199
they may find out that because it's
1537
01:03:59,919 --> 01:04:05,069
warmer in the afternoon that you need to
1538
01:04:02,199 --> 01:04:07,869
adjust the humidity and temperature
1539
01:04:05,069 --> 01:04:09,849
depending on the time of day based on
1540
01:04:07,869 --> 01:04:12,400
the insights from your data science team
1541
01:04:09,849 --> 01:04:15,130
you get suggestions for hypotheses and
1542
01:04:12,400 --> 01:04:17,289
actions on how to change the operations
1543
01:04:15,130 --> 01:04:19,809
of the manufacturing line in order to
1544
01:04:17,289 --> 01:04:21,999
improve the productivity of the line and
1545
01:04:19,809 --> 01:04:24,039
when you deploy the changes you then get
1546
01:04:21,999 --> 01:04:26,019
new data back that you can be analyzed
1547
01:04:24,039 --> 01:04:28,119
periodically so you can keep on
1548
01:04:26,019 --> 01:04:31,119
optimizing the performance of your
1549
01:04:28,119 --> 01:04:32,859
manufacturing line to summarize the key
1550
01:04:31,119 --> 01:04:35,709
steps of a data science project are to
1551
01:04:32,859 --> 01:04:38,859
collect the data to analyze the data and
1552
01:04:35,709 --> 01:04:41,229
then to suggest hypotheses and actions
1553
01:04:38,859 --> 01:04:43,390
in this video in the last video you saw
1554
01:04:41,229 --> 01:04:45,939
some examples of machine learning
1555
01:04:43,390 --> 01:04:47,919
projects and data science projects it
1556
01:04:45,939 --> 01:04:50,259
turns out that machine learning and data
1557
01:04:47,919 --> 01:04:52,869
science are affecting almost every
1558
01:04:50,259 --> 01:04:54,819
single job function what I want to do in
1559
01:04:52,869 --> 01:04:57,279
the next video is show you how these
1560
01:04:54,819 --> 01:04:59,919
ideas are affecting many job functions
1561
01:04:57,279 --> 01:05:02,199
including perhaps yours and certainly
1562
01:04:59,919 --> 01:05:03,320
that many of your colleagues let's go
1563
01:05:02,199 --> 01:05:05,380
into the next video
1564
01:05:03,320 --> 01:05:05,380
you
1565
01:05:07,230 --> 01:05:13,020
data is transforming many different job
1566
01:05:10,380 --> 01:05:15,240
functions whether you work in recruiting
1567
01:05:13,020 --> 01:05:18,109
or sales or marketing or manufacturing
1568
01:05:15,240 --> 01:05:20,760
or agriculture data is probably
1569
01:05:18,109 --> 01:05:22,470
transforming your job function what's
1570
01:05:20,760 --> 01:05:25,079
happened in the last few decades is the
1571
01:05:22,470 --> 01:05:26,640
digitization of our society so rather
1572
01:05:25,079 --> 01:05:29,490
than handing out paper surveys like
1573
01:05:26,640 --> 01:05:32,490
these surveys are more likely to be done
1574
01:05:29,490 --> 01:05:34,349
in digital format or doctors still write
1575
01:05:32,490 --> 01:05:36,150
some handwritten notes but the doctors
1576
01:05:34,349 --> 01:05:39,150
handwritten note is increasingly likely
1577
01:05:36,150 --> 01:05:41,549
to be a digital record and so to this in
1578
01:05:39,150 --> 01:05:44,040
just about every single job function and
1579
01:05:41,549 --> 01:05:45,599
this availability of data means that
1580
01:05:44,040 --> 01:05:48,240
there's a good chance that your job
1581
01:05:45,599 --> 01:05:50,520
function can be helped with tools like
1582
01:05:48,240 --> 01:05:52,200
data science or machine learning let's
1583
01:05:50,520 --> 01:05:53,940
take a look and in this video I want to
1584
01:05:52,200 --> 01:05:56,490
run through many different job functions
1585
01:05:53,940 --> 01:05:59,220
and discuss how data science and machine
1586
01:05:56,490 --> 01:06:02,220
learning can or will impact these
1587
01:05:59,220 --> 01:06:05,130
different types of jobs there's lots of
1588
01:06:02,220 --> 01:06:07,410
sales you've already seen in the last
1589
01:06:05,130 --> 01:06:09,180
video how data science can be used to
1590
01:06:07,410 --> 01:06:11,400
optimize a sales funnel
1591
01:06:09,180 --> 01:06:13,410
how about machine learning if you're a
1592
01:06:11,400 --> 01:06:15,329
salesperson you may have a set of leaves
1593
01:06:13,410 --> 01:06:17,790
about different people that you can
1594
01:06:15,329 --> 01:06:20,130
reach out to to convince them to buy
1595
01:06:17,790 --> 01:06:22,710
something from your company machine
1596
01:06:20,130 --> 01:06:25,200
learning can help you prioritize these
1597
01:06:22,710 --> 01:06:28,020
leads so you might want to prioritize
1598
01:06:25,200 --> 01:06:29,880
calling out the CEO of the large company
1599
01:06:28,020 --> 01:06:33,059
rather than the intern at a much smaller
1600
01:06:29,880 --> 01:06:35,339
company and this type of automated lead
1601
01:06:33,059 --> 01:06:37,079
sorting is making salespeople more
1602
01:06:35,339 --> 01:06:39,390
efficient let's look at more examples
1603
01:06:37,079 --> 01:06:41,790
let's say your manufacturing line
1604
01:06:39,390 --> 01:06:43,859
manager you've already seen how data
1605
01:06:41,790 --> 01:06:44,730
science can help you optimize a
1606
01:06:43,859 --> 01:06:47,220
manufacturing line
1607
01:06:44,730 --> 01:06:49,230
how about machine learning one of the
1608
01:06:47,220 --> 01:06:51,510
steps of this manufacturing process is
1609
01:06:49,230 --> 01:06:54,000
the final inspection and in fact today
1610
01:06:51,510 --> 01:06:56,160
in many factories there can be hundreds
1611
01:06:54,000 --> 01:06:59,160
or thousands of people using the human
1612
01:06:56,160 --> 01:07:01,290
eye to check over objects maybe coffee
1613
01:06:59,160 --> 01:07:02,910
mugs maybe other things to see if
1614
01:07:01,290 --> 01:07:05,609
there's scratches or dents and that's
1615
01:07:02,910 --> 01:07:09,059
called inspection so machine learning
1616
01:07:05,609 --> 01:07:12,420
can take us input data set like this and
1617
01:07:09,059 --> 01:07:15,480
learn to automatically figure out if a
1618
01:07:12,420 --> 01:07:18,690
coffee mug is defective or not and by
1619
01:07:15,480 --> 01:07:20,020
automatically finding scratches or dents
1620
01:07:18,690 --> 01:07:22,270
it can
1621
01:07:20,020 --> 01:07:25,360
reduce labor costs and also improve
1622
01:07:22,270 --> 01:07:27,340
quality in your factory this type of
1623
01:07:25,360 --> 01:07:28,870
automated visual inspection is one of
1624
01:07:27,340 --> 01:07:31,240
the technologies that I think will have
1625
01:07:28,870 --> 01:07:33,070
a big impact on manufacturing this is
1626
01:07:31,240 --> 01:07:35,380
something I've been working on myself as
1627
01:07:33,070 --> 01:07:37,810
well let's see more examples how about
1628
01:07:35,380 --> 01:07:39,640
recruiting when recruiting someone to
1629
01:07:37,810 --> 01:07:42,220
join your company there may be a pretty
1630
01:07:39,640 --> 01:07:44,650
predictable sequence of steps where your
1631
01:07:42,220 --> 01:07:47,290
recruiter or someone else would send an
1632
01:07:44,650 --> 01:07:49,540
email to a candidate and then you'd have
1633
01:07:47,290 --> 01:07:51,520
a phone call of them bring them on-site
1634
01:07:49,540 --> 01:07:54,130
for an interview and then extend an
1635
01:07:51,520 --> 01:07:56,410
offer and maybe close the offer similar
1636
01:07:54,130 --> 01:08:00,280
to how data science can be used to
1637
01:07:56,410 --> 01:08:02,440
optimize a sales funnel recruiting can
1638
01:08:00,280 --> 01:08:04,510
also use data science to optimize a
1639
01:08:02,440 --> 01:08:06,670
recruiting funnel and in fact many
1640
01:08:04,510 --> 01:08:09,430
recruiting organizations are doing so
1641
01:08:06,670 --> 01:08:11,560
today for example if you find that
1642
01:08:09,430 --> 01:08:13,090
hardly anyone is making it from the
1643
01:08:11,560 --> 01:08:15,100
phone screen step to the on-site
1644
01:08:13,090 --> 01:08:17,529
interview step then you may conclude
1645
01:08:15,100 --> 01:08:19,420
that maybe too many people are getting
1646
01:08:17,529 --> 01:08:21,069
to the phone screen stage or maybe the
1647
01:08:19,420 --> 01:08:22,510
people doing the phone screen are just
1648
01:08:21,069 --> 01:08:24,430
being too tough and they should let more
1649
01:08:22,510 --> 01:08:26,589
people get to the on-site interview
1650
01:08:24,430 --> 01:08:29,080
stage this type of data science is
1651
01:08:26,589 --> 01:08:31,000
already having an impact on recruiting
1652
01:08:29,080 --> 01:08:33,609
what about machine learning projects
1653
01:08:31,000 --> 01:08:36,190
well one of the steps of recruiting is
1654
01:08:33,609 --> 01:08:38,529
to screen a lot of resumes to decide who
1655
01:08:36,190 --> 01:08:41,049
to reach out to so you may have to look
1656
01:08:38,529 --> 01:08:42,730
at my resume and says yes let's email
1657
01:08:41,049 --> 01:08:44,890
them don't get a different one to say no
1658
01:08:42,730 --> 01:08:47,230
let's not move ahead with this candidate
1659
01:08:44,890 --> 01:08:49,960
and machine learning is starting to make
1660
01:08:47,230 --> 01:08:52,029
us weigh into automated resume screening
1661
01:08:49,960 --> 01:08:54,400
this does raise important ethical
1662
01:08:52,029 --> 01:08:57,250
questions such as making sure that your
1663
01:08:54,400 --> 01:08:59,560
AI software does not exhibit undesirable
1664
01:08:57,250 --> 01:09:01,690
forms of bias and treat people fairly
1665
01:08:59,560 --> 01:09:04,000
but machine learning is starting to make
1666
01:09:01,690 --> 01:09:05,950
inroads into this and hope can do so
1667
01:09:04,000 --> 01:09:08,350
while making sure that the systems are
1668
01:09:05,950 --> 01:09:10,569
ethical and fair in the final week of
1669
01:09:08,350 --> 01:09:12,970
this AI for everyone calls you also
1670
01:09:10,569 --> 01:09:16,029
learn more about the issues of fairness
1671
01:09:12,970 --> 01:09:18,040
and ethics in AI wonder if you work in
1672
01:09:16,029 --> 01:09:20,020
marketing one of the common ways to
1673
01:09:18,040 --> 01:09:22,510
optimize that performance on the website
1674
01:09:20,020 --> 01:09:24,730
is called a be testing in which you
1675
01:09:22,510 --> 01:09:27,040
launch two versions of website here in
1676
01:09:24,730 --> 01:09:29,170
version 8 has a red button version B has
1677
01:09:27,040 --> 01:09:32,249
a green button and you'd measure which
1678
01:09:29,170 --> 01:09:35,099
website causes people to click through
1679
01:09:32,249 --> 01:09:37,380
so with this type of data a data science
1680
01:09:35,099 --> 01:09:39,599
team can help you gain insights and
1681
01:09:37,380 --> 01:09:42,089
suggest hypotheses or actions for
1682
01:09:39,599 --> 01:09:44,549
optimizing a website how about machine
1683
01:09:42,089 --> 01:09:46,799
learning and marketing today a lot of
1684
01:09:44,549 --> 01:09:48,539
websites will give customized product
1685
01:09:46,799 --> 01:09:50,309
recommendations to show you the things
1686
01:09:48,539 --> 01:09:52,229
you are most likely to want to buy and
1687
01:09:50,309 --> 01:09:54,539
there's actually significant increases
1688
01:09:52,229 --> 01:09:56,699
sales on these websites for example a
1689
01:09:54,539 --> 01:09:59,070
clothing website after has seen the way
1690
01:09:56,699 --> 01:10:00,659
I shop after a while will hopefully just
1691
01:09:59,070 --> 01:10:02,579
recommend blue shirts to me because
1692
01:10:00,659 --> 01:10:04,829
that's frankly pretty much the only type
1693
01:10:02,579 --> 01:10:06,780
of shirt I ever buy but maybe other
1694
01:10:04,829 --> 01:10:08,369
customers will have more diverse and
1695
01:10:06,780 --> 01:10:11,099
more interesting recommendations than
1696
01:10:08,369 --> 01:10:13,199
mine but today these customized product
1697
01:10:11,099 --> 01:10:15,599
recommendations actually drive a large
1698
01:10:13,199 --> 01:10:18,360
percentage of sales on many large online
1699
01:10:15,599 --> 01:10:20,849
e-commerce websites one last example
1700
01:10:18,360 --> 01:10:22,889
from a totally different sector let's
1701
01:10:20,849 --> 01:10:24,659
say you work in agriculture maybe you're
1702
01:10:22,889 --> 01:10:27,630
a farmer working on the large industrial
1703
01:10:24,659 --> 01:10:30,269
farm how can data science help you today
1704
01:10:27,630 --> 01:10:32,610
farmers already using data signs for
1705
01:10:30,269 --> 01:10:34,559
crop analytics where you can take data
1706
01:10:32,610 --> 01:10:36,630
on the soil conditions the weather
1707
01:10:34,559 --> 01:10:38,729
conditions the prices of different crops
1708
01:10:36,630 --> 01:10:41,249
in the market and have data science
1709
01:10:38,729 --> 01:10:44,099
teams make recommendations to what to
1710
01:10:41,249 --> 01:10:46,530
plant when to plant so as to improve use
1711
01:10:44,099 --> 01:10:47,639
while maintaining the condition of the
1712
01:10:46,530 --> 01:10:50,099
soil on your farm
1713
01:10:47,639 --> 01:10:51,719
this type of data science is and will
1714
01:10:50,099 --> 01:10:53,940
play a bigger and bigger role in
1715
01:10:51,719 --> 01:10:56,159
agriculture let's also look at the
1716
01:10:53,940 --> 01:10:58,710
machine learning example I think one of
1717
01:10:56,159 --> 01:11:01,199
the most exciting changes to agriculture
1718
01:10:58,710 --> 01:11:03,539
is precision agriculture here's a
1719
01:11:01,199 --> 01:11:05,489
picture that I took on a farm with my
1720
01:11:03,539 --> 01:11:08,099
cell phone on the upper right is a
1721
01:11:05,489 --> 01:11:11,010
cotton plant and shown in middle is a
1722
01:11:08,099 --> 01:11:12,869
weed and so with machine learning we're
1723
01:11:11,010 --> 01:11:14,429
starting to see products that can go on
1724
01:11:12,869 --> 01:11:18,599
to the farms take a picture like this
1725
01:11:14,429 --> 01:11:20,729
and spray a read color in a very precise
1726
01:11:18,599 --> 01:11:22,499
way just onto the weeds so that it gets
1727
01:11:20,729 --> 01:11:25,019
your the read but without having to
1728
01:11:22,499 --> 01:11:27,329
spray an excessive amount of read colors
1729
01:11:25,019 --> 01:11:30,210
this type of machine learning technology
1730
01:11:27,329 --> 01:11:33,059
is both helping farmers increased crop
1731
01:11:30,210 --> 01:11:35,400
use while also hoping to preserve the
1732
01:11:33,059 --> 01:11:37,499
environment in this video you saw how
1733
01:11:35,400 --> 01:11:39,929
all of these job functions everything
1734
01:11:37,499 --> 01:11:42,239
from sales recruiting to marketing to
1735
01:11:39,929 --> 01:11:44,400
manufacturing to farming agriculture how
1736
01:11:42,239 --> 01:11:45,150
all of these job functions are being
1737
01:11:44,400 --> 01:11:46,860
affected
1738
01:11:45,150 --> 01:11:49,230
data by data science and the machine
1739
01:11:46,860 --> 01:11:50,850
learning it seems like there's a lot of
1740
01:11:49,230 --> 01:11:53,159
different things you could do with AI
1741
01:11:50,850 --> 01:11:55,199
but how do you actually select a
1742
01:11:53,159 --> 01:11:58,699
promising project to work on let's talk
1743
01:11:55,199 --> 01:11:58,699
about that in the next video
1744
01:12:01,460 --> 01:12:07,400
if you want to try your hand at an AI
1745
01:12:04,760 --> 01:12:09,830
project how do you select a worthwhile
1746
01:12:07,400 --> 01:12:11,929
project to work on don't expect an idea
1747
01:12:09,830 --> 01:12:13,640
it's an S we come overnight sometimes it
1748
01:12:11,929 --> 01:12:15,920
happens but sometimes it also takes a
1749
01:12:13,640 --> 01:12:18,260
few days or maybe a few weeks to come up
1750
01:12:15,920 --> 01:12:19,820
with a worthy idea to pursue in this
1751
01:12:18,260 --> 01:12:22,130
video you see a framework for
1752
01:12:19,820 --> 01:12:24,830
brainstorming potentially exciting the
1753
01:12:22,130 --> 01:12:26,929
AI projects to pursue let's say you want
1754
01:12:24,830 --> 01:12:29,060
to build an AI project for your business
1755
01:12:26,929 --> 01:12:31,730
you've already seen that AI can't do
1756
01:12:29,060 --> 01:12:34,810
everything and so there's going to be a
1757
01:12:31,730 --> 01:12:37,880
certain set of things that is what AI
1758
01:12:34,810 --> 01:12:40,670
can do so let's let the circle represent
1759
01:12:37,880 --> 01:12:43,010
a set of things that AI can do now
1760
01:12:40,670 --> 01:12:45,890
there's also going to be a certain set
1761
01:12:43,010 --> 01:12:48,800
of things that is valuable for your
1762
01:12:45,890 --> 01:12:51,140
business so let's let this second circle
1763
01:12:48,800 --> 01:12:53,239
represent a set of things that are
1764
01:12:51,140 --> 01:12:55,100
valuable for your business what you
1765
01:12:53,239 --> 01:12:57,679
would like to do is try to select
1766
01:12:55,100 --> 01:13:00,770
projects that are at the intersection of
1767
01:12:57,679 --> 01:13:02,570
these two sets so you select projects
1768
01:13:00,770 --> 01:13:04,610
hopefully that are both feasible that
1769
01:13:02,570 --> 01:13:07,850
can be done with AI and that are also
1770
01:13:04,610 --> 01:13:09,890
valuable for your business so AI experts
1771
01:13:07,850 --> 01:13:12,110
will tend to have a good sense of what
1772
01:13:09,890 --> 01:13:14,690
is and what isn't in the set on the left
1773
01:13:12,110 --> 01:13:16,400
and domain expense expense in your
1774
01:13:14,690 --> 01:13:18,290
business be it sales or marketing or
1775
01:13:16,400 --> 01:13:19,820
agriculture or something else what have
1776
01:13:18,290 --> 01:13:22,130
a best sense of what is actually
1777
01:13:19,820 --> 01:13:24,710
valuable for your business so when
1778
01:13:22,130 --> 01:13:26,900
brainstorming projects that AI can do
1779
01:13:24,710 --> 01:13:29,270
and are valid for your business I will
1780
01:13:26,900 --> 01:13:32,060
often bring together a team comprising
1781
01:13:29,270 --> 01:13:34,909
both people knowledgeable AI as well as
1782
01:13:32,060 --> 01:13:37,040
experts in your business area to
1783
01:13:34,909 --> 01:13:39,409
brainstorm together so that together
1784
01:13:37,040 --> 01:13:41,659
they can try to identify projects at the
1785
01:13:39,409 --> 01:13:44,409
intersection of both of these two sets
1786
01:13:41,659 --> 01:13:46,489
so sometimes we also call these
1787
01:13:44,409 --> 01:13:48,980
cross-functional teams and that just
1788
01:13:46,489 --> 01:13:51,739
means a team that includes both AI
1789
01:13:48,980 --> 01:13:53,870
experts as well as domain experts
1790
01:13:51,739 --> 01:13:57,110
meaning experts in your area of business
1791
01:13:53,870 --> 01:13:59,390
when brainstorming projects there's a
1792
01:13:57,110 --> 01:14:02,420
framework that I've used with a lot of
1793
01:13:59,390 --> 01:14:05,120
companies that are found to be useful so
1794
01:14:02,420 --> 01:14:07,670
let me share with you three principles
1795
01:14:05,120 --> 01:14:10,909
or three ideas for how you can have a
1796
01:14:07,670 --> 01:14:12,170
team brainstorm projects first even
1797
01:14:10,909 --> 01:14:15,209
though there's been a lot of press
1798
01:14:12,170 --> 01:14:17,670
coverage about AI automating jobs away
1799
01:14:15,209 --> 01:14:19,829
this is an important societal issue that
1800
01:14:17,670 --> 01:14:22,170
needs to be addressed when thinking
1801
01:14:19,829 --> 01:14:23,429
about concrete AI projects I find it
1802
01:14:22,170 --> 01:14:26,939
much more useful to think about
1803
01:14:23,429 --> 01:14:30,929
automating tasks rather than automating
1804
01:14:26,939 --> 01:14:32,639
jobs take call center operations there
1805
01:14:30,929 --> 01:14:34,829
longer tasks that happen in the call
1806
01:14:32,639 --> 01:14:36,360
center ranging from people pick up the
1807
01:14:34,829 --> 01:14:38,729
phone to answering phone calls to
1808
01:14:36,360 --> 01:14:40,860
replying to emails to taking specific
1809
01:14:38,729 --> 01:14:43,079
actions such as issuing a refund on
1810
01:14:40,860 --> 01:14:44,939
behalf of a customer requests but along
1811
01:14:43,079 --> 01:14:47,849
with these tasks that employees in the
1812
01:14:44,939 --> 01:14:49,979
call center do there may be one call
1813
01:14:47,849 --> 01:14:51,749
routing or email routing that may be
1814
01:14:49,979 --> 01:14:53,280
particularly amenable so machine
1815
01:14:51,749 --> 01:14:55,019
learning automation and it's been
1816
01:14:53,280 --> 01:14:57,329
looking at all these tasks that the
1817
01:14:55,019 --> 01:14:59,550
group of employees do and selecting one
1818
01:14:57,329 --> 01:15:01,979
that we allow you to select the most
1819
01:14:59,550 --> 01:15:04,199
fruitful project for automation in the
1820
01:15:01,979 --> 01:15:06,869
near term let's look at another example
1821
01:15:04,199 --> 01:15:10,159
the job of radiologists there's no
1822
01:15:06,869 --> 01:15:13,229
longer press about how a I'm a automate
1823
01:15:10,159 --> 01:15:14,849
radiologist jobs but radiologists
1824
01:15:13,229 --> 01:15:17,429
actually do a lot of things they read
1825
01:15:14,849 --> 01:15:19,469
x-rays that's really important but they
1826
01:15:17,429 --> 01:15:21,719
also engage in their own continuing
1827
01:15:19,469 --> 01:15:24,119
education they consulted other doctors
1828
01:15:21,719 --> 01:15:26,400
they may mentor younger doctors some of
1829
01:15:24,119 --> 01:15:28,709
them also consult directly with patients
1830
01:15:26,400 --> 01:15:31,199
and so it's by looking at all of these
1831
01:15:28,709 --> 01:15:34,170
tasks that the radiologist does that you
1832
01:15:31,199 --> 01:15:36,989
may identify one of them let's say AI
1833
01:15:34,170 --> 01:15:39,030
assistants or AI automation for reading
1834
01:15:36,989 --> 01:15:41,639
x-rays that allows you to select the
1835
01:15:39,030 --> 01:15:43,619
most fruitful projects to work on so
1836
01:15:41,639 --> 01:15:44,340
whether we recommend is if you look in
1837
01:15:43,619 --> 01:15:46,679
your business
1838
01:15:44,340 --> 01:15:48,869
think about the tasks that people do to
1839
01:15:46,679 --> 01:15:50,729
see if you can identify just one of them
1840
01:15:48,869 --> 01:15:53,849
or just a couple of them that may be
1841
01:15:50,729 --> 01:15:56,429
automatable using machine learning when
1842
01:15:53,849 --> 01:15:58,979
our meeting CEOs of large companies to
1843
01:15:56,429 --> 01:16:01,800
brainstorm AI projects for the company a
1844
01:15:58,979 --> 01:16:03,739
common question out also ask is what are
1845
01:16:01,800 --> 01:16:06,449
the main drivers of business value and
1846
01:16:03,739 --> 01:16:08,939
sometimes finding AI solution so they
1847
01:16:06,449 --> 01:16:11,880
design solutions to augmentis can be
1848
01:16:08,939 --> 01:16:13,650
very valuable finally a third question
1849
01:16:11,880 --> 01:16:16,409
that I've asked there's sometimes letter
1850
01:16:13,650 --> 01:16:18,539
valuable project ideas is what the main
1851
01:16:16,409 --> 01:16:20,489
pain points in your business some of
1852
01:16:18,539 --> 01:16:22,349
them could be soft of AI some of them
1853
01:16:20,489 --> 01:16:24,150
can't be soft for the eye but by
1854
01:16:22,349 --> 01:16:26,010
understanding the main pain points of
1855
01:16:24,150 --> 01:16:28,409
the business that can create a useful
1856
01:16:26,010 --> 01:16:31,409
starting point for brainstorming AI
1857
01:16:28,409 --> 01:16:34,289
projects as well I have one last piece
1858
01:16:31,409 --> 01:16:36,599
of advice for brainstorming AI projects
1859
01:16:34,289 --> 01:16:38,130
which is that you can make progress even
1860
01:16:36,599 --> 01:16:40,979
without big data
1861
01:16:38,130 --> 01:16:43,320
even without tons of data now don't get
1862
01:16:40,979 --> 01:16:45,599
me wrong having more data almost never
1863
01:16:43,320 --> 01:16:47,340
hurts other than maybe needing to pay a
1864
01:16:45,599 --> 01:16:49,409
bit more for disk space or network
1865
01:16:47,340 --> 01:16:52,650
bandwidth to transmit and store the data
1866
01:16:49,409 --> 01:16:54,719
having more data almost always is only
1867
01:16:52,650 --> 01:16:57,570
helpful and I love having lots of data
1868
01:16:54,719 --> 01:16:59,969
it is also true that data makes some
1869
01:16:57,570 --> 01:17:02,280
businesses like web search defensible
1870
01:16:59,969 --> 01:17:04,079
web search is a longtail business
1871
01:17:02,280 --> 01:17:07,260
meaning that there are a lot of very
1872
01:17:04,079 --> 01:17:09,420
very rare web queries and so seeing what
1873
01:17:07,260 --> 01:17:12,239
people click on when they search on all
1874
01:17:09,420 --> 01:17:14,579
of these rare web queries does help the
1875
01:17:12,239 --> 01:17:18,269
leading web search engines have a much
1876
01:17:14,579 --> 01:17:20,849
better search experience so big data is
1877
01:17:18,269 --> 01:17:23,249
great when you can get it but I think
1878
01:17:20,849 --> 01:17:26,130
big data is also sometimes overhyped and
1879
01:17:23,249 --> 01:17:28,769
even with a small data set you can still
1880
01:17:26,130 --> 01:17:31,709
often make progress here's an example
1881
01:17:28,769 --> 01:17:33,749
let's say you're building a automated
1882
01:17:31,709 --> 01:17:35,670
visual inspection system for the coffee
1883
01:17:33,749 --> 01:17:37,800
mug so you want to automatically detect
1884
01:17:35,670 --> 01:17:40,650
that the coffee mug on the right is
1885
01:17:37,800 --> 01:17:42,780
defective well if you had a million
1886
01:17:40,650 --> 01:17:44,699
pictures of good coffee mugs and
1887
01:17:42,780 --> 01:17:47,099
defective coffee mugs it'd be great to
1888
01:17:44,699 --> 01:17:48,989
have that many examples of pictures of
1889
01:17:47,099 --> 01:17:51,630
good and bad coffee mouths to feed
1890
01:17:48,989 --> 01:17:53,369
through AI system but I hope that you
1891
01:17:51,630 --> 01:17:55,019
have not manufactured a million
1892
01:17:53,369 --> 01:17:56,849
defective coffee mouths because that
1893
01:17:55,019 --> 01:17:59,999
feels like a very expensive thing to
1894
01:17:56,849 --> 01:18:03,510
have to throw away so sometimes with as
1895
01:17:59,999 --> 01:18:05,849
few as a hundred pictures or maybe a
1896
01:18:03,510 --> 01:18:08,729
thousand or sometimes maybe as few as
1897
01:18:05,849 --> 01:18:10,949
ten you may be able to get started on a
1898
01:18:08,729 --> 01:18:13,469
machine learning project the amount of
1899
01:18:10,949 --> 01:18:16,409
data you need is very problem dependent
1900
01:18:13,469 --> 01:18:18,689
and speaking with a ai engineer on AI
1901
01:18:16,409 --> 01:18:21,090
expert would help you get better since
1902
01:18:18,689 --> 01:18:23,219
there are some problems for a thousand
1903
01:18:21,090 --> 01:18:25,079
images may not be enough where you do
1904
01:18:23,219 --> 01:18:28,110
need big data to get good performance
1905
01:18:25,079 --> 01:18:30,090
but my advice is don't give up just
1906
01:18:28,110 --> 01:18:32,249
because you don't have a lot of data to
1907
01:18:30,090 --> 01:18:35,130
start off with and you can often still
1908
01:18:32,249 --> 01:18:37,380
make progress even with a small data set
1909
01:18:35,130 --> 01:18:39,600
in this video you saw a brainstorming
1910
01:18:37,380 --> 01:18:41,700
framework and a set of criteria for
1911
01:18:39,600 --> 01:18:44,790
trying to come up with projects that
1912
01:18:41,700 --> 01:18:47,160
hopefully can be doable with AI and
1913
01:18:44,790 --> 01:18:50,100
they're also valuable for your business
1914
01:18:47,160 --> 01:18:52,440
now having brainstormed elicit projects
1915
01:18:50,100 --> 01:18:54,480
how do you select one or select a small
1916
01:18:52,440 --> 01:18:56,760
handful to actually commit to and work
1917
01:18:54,480 --> 01:18:58,940
on let's talk about that in the next
1918
01:18:56,760 --> 01:18:58,940
video
1919
01:19:01,690 --> 01:19:06,280
maybe have a lot of ideas for possible
1920
01:19:04,270 --> 01:19:08,320
AI projects to work on
1921
01:19:06,280 --> 01:19:10,000
but before committing to one how do you
1922
01:19:08,320 --> 01:19:13,180
make sure that this really is a
1923
01:19:10,000 --> 01:19:14,620
worthwhile project if is a quick project
1924
01:19:13,180 --> 01:19:16,540
that might take you just a few days
1925
01:19:14,620 --> 01:19:18,730
maybe just jump in right away and see
1926
01:19:16,540 --> 01:19:21,699
the worse or not but some AI projects
1927
01:19:18,730 --> 01:19:23,199
may take many months to execute in this
1928
01:19:21,699 --> 01:19:25,900
video I want to step you through the
1929
01:19:23,199 --> 01:19:27,910
process that I use to double check if a
1930
01:19:25,900 --> 01:19:30,130
project is worth that many months of
1931
01:19:27,910 --> 01:19:32,710
effort let's take a look before
1932
01:19:30,130 --> 01:19:36,250
committing to a big AI project I will
1933
01:19:32,710 --> 01:19:39,160
usually conduct due diligence on it due
1934
01:19:36,250 --> 01:19:42,280
diligence has a specific meaning in the
1935
01:19:39,160 --> 01:19:44,080
legal world but informally it just means
1936
01:19:42,280 --> 01:19:46,300
that you want to spend some time to make
1937
01:19:44,080 --> 01:19:49,660
sure what your hope is true really is
1938
01:19:46,300 --> 01:19:52,270
true you've already seen how the best AI
1939
01:19:49,660 --> 01:19:54,820
projects are ones that are feasible so
1940
01:19:52,270 --> 01:19:57,370
it's something that a I can do as well
1941
01:19:54,820 --> 01:19:59,580
as valuable we really want to choose
1942
01:19:57,370 --> 01:20:04,480
projects to the at the intersection of
1943
01:19:59,580 --> 01:20:07,449
these two sets so to make sure a project
1944
01:20:04,480 --> 01:20:10,840
is feasible I will usually go through
1945
01:20:07,449 --> 01:20:13,449
technical diligence and to make sure
1946
01:20:10,840 --> 01:20:16,150
that the project is valuable I will
1947
01:20:13,449 --> 01:20:18,160
usually go through a business diligence
1948
01:20:16,150 --> 01:20:20,890
process let me tell you more about these
1949
01:20:18,160 --> 01:20:23,739
two steps technical diligence is the
1950
01:20:20,890 --> 01:20:25,900
process of making sure that the AI
1951
01:20:23,739 --> 01:20:30,250
system you hope to build really is
1952
01:20:25,900 --> 01:20:32,949
doable really is feasible so you might
1953
01:20:30,250 --> 01:20:35,290
talk to AI X phase about whether or not
1954
01:20:32,949 --> 01:20:37,420
the AI system can actually meet the
1955
01:20:35,290 --> 01:20:40,150
desired level of performance for example
1956
01:20:37,420 --> 01:20:43,930
if you are hoping to build a speech
1957
01:20:40,150 --> 01:20:46,390
system that is 95% accurate Consulting
1958
01:20:43,930 --> 01:20:48,550
of AI experience or perhaps reading some
1959
01:20:46,390 --> 01:20:50,560
of the trade literature can give you a
1960
01:20:48,550 --> 01:20:54,190
sense of whether this is doable or not
1961
01:20:50,560 --> 01:20:56,440
or if you want a system to inspect
1962
01:20:54,190 --> 01:20:59,949
coffee mugs in the factory and you need
1963
01:20:56,440 --> 01:21:01,600
your system to be 99% accurate again is
1964
01:20:59,949 --> 01:21:04,000
this actually doable with today's
1965
01:21:01,600 --> 01:21:06,430
technology a second important question
1966
01:21:04,000 --> 01:21:09,160
for technical diligence is how much data
1967
01:21:06,430 --> 01:21:11,739
is needed to get to this design level
1968
01:21:09,160 --> 01:21:12,869
performance and do you have a way to get
1969
01:21:11,739 --> 01:21:15,999
that much
1970
01:21:12,869 --> 01:21:17,499
third would be engineering timeline to
1971
01:21:15,999 --> 01:21:20,170
try to figure out how long it will take
1972
01:21:17,499 --> 01:21:22,119
and how many people will take to build
1973
01:21:20,170 --> 01:21:24,550
the system that you would like to have
1974
01:21:22,119 --> 01:21:27,280
built in addition to technical divisions
1975
01:21:24,550 --> 01:21:29,469
I will often also conduct business
1976
01:21:27,280 --> 01:21:32,110
diligence to make sure that the project
1977
01:21:29,469 --> 01:21:35,499
you envision really is valuable for the
1978
01:21:32,110 --> 01:21:38,050
business so a lot of AI projects would
1979
01:21:35,499 --> 01:21:41,739
drive value through lowering costs for
1980
01:21:38,050 --> 01:21:44,289
example by automating a few tasks or by
1981
01:21:41,739 --> 01:21:47,320
squeezing more efficiency out of a
1982
01:21:44,289 --> 01:21:50,199
system a lot of AI systems can also
1983
01:21:47,320 --> 01:21:52,539
increase revenue for example driving
1984
01:21:50,199 --> 01:21:55,030
more people to check out in your
1985
01:21:52,539 --> 01:21:57,099
shopping cart or you may be building an
1986
01:21:55,030 --> 01:22:00,010
AI system to help you launch a new
1987
01:21:57,099 --> 01:22:02,260
product or a new line of business so
1988
01:22:00,010 --> 01:22:04,630
business diligence is the process of
1989
01:22:02,260 --> 01:22:06,670
thinking through carefully for the AI
1990
01:22:04,630 --> 01:22:08,650
system that you're building such as a
1991
01:22:06,670 --> 01:22:11,469
speech recognition system that's 95%
1992
01:22:08,650 --> 01:22:14,980
accurate or a visual inspection system
1993
01:22:11,469 --> 01:22:17,679
does 99.9% accurate would allow you to
1994
01:22:14,980 --> 01:22:20,499
achieve your business goals whether your
1995
01:22:17,679 --> 01:22:23,800
business goal is to improve your current
1996
01:22:20,499 --> 01:22:25,809
business or to even create brand new
1997
01:22:23,800 --> 01:22:27,789
businesses in your company when
1998
01:22:25,809 --> 01:22:30,309
conducting business diligence I'll often
1999
01:22:27,789 --> 01:22:31,989
end up building spreadsheet financial
2000
01:22:30,309 --> 01:22:35,019
models to estimate the value
2001
01:22:31,989 --> 01:22:37,659
quantitatively such as estimate how many
2002
01:22:35,019 --> 01:22:39,400
dollars are actually saved or what do we
2003
01:22:37,659 --> 01:22:41,949
think is a reasonable assumption in
2004
01:22:39,400 --> 01:22:44,199
terms of increase revenue and to model
2005
01:22:41,949 --> 01:22:46,929
out the economics associated with a
2006
01:22:44,199 --> 01:22:49,030
project before committing to many months
2007
01:22:46,929 --> 01:22:50,590
of effort on the project although not
2008
01:22:49,030 --> 01:22:52,900
explicitly listed on this slide
2009
01:22:50,590 --> 01:22:54,820
one thing I hope you're also considering
2010
01:22:52,900 --> 01:22:57,369
is the third type of diligence which is
2011
01:22:54,820 --> 01:23:00,159
ethical diligence I think there are a
2012
01:22:57,369 --> 01:23:02,199
lot of things that AI can do that will
2013
01:23:00,159 --> 01:23:04,539
even make a lot of money but that may
2014
01:23:02,199 --> 01:23:07,059
not make society better off so in
2015
01:23:04,539 --> 01:23:09,039
addition to technical diligence and
2016
01:23:07,059 --> 01:23:11,320
business diligence I hope you also
2017
01:23:09,039 --> 01:23:12,849
conduct ethical diligence and make sure
2018
01:23:11,320 --> 01:23:15,550
that what are you doing is actually
2019
01:23:12,849 --> 01:23:17,860
making humanity and making society
2020
01:23:15,550 --> 01:23:20,079
better off we also talked more about
2021
01:23:17,860 --> 01:23:22,900
this in the last week of this course as
2022
01:23:20,079 --> 01:23:25,479
well as you're planning out your AI
2023
01:23:22,900 --> 01:23:26,350
project you also have to decide do you
2024
01:23:25,479 --> 01:23:29,620
want to
2025
01:23:26,350 --> 01:23:32,170
or by this is an age-old question in the
2026
01:23:29,620 --> 01:23:35,380
IT world and we're facing this question
2027
01:23:32,170 --> 01:23:37,570
in AI as well for example hardly any
2028
01:23:35,380 --> 01:23:39,400
companies built their own computers
2029
01:23:37,570 --> 01:23:42,220
these days they buy someone else's
2030
01:23:39,400 --> 01:23:45,010
computers and hardly any companies build
2031
01:23:42,220 --> 01:23:47,110
their own Wi-Fi routers just by a
2032
01:23:45,010 --> 01:23:48,850
commercial Wi-Fi router
2033
01:23:47,110 --> 01:23:51,280
how about machine learning and data
2034
01:23:48,850 --> 01:23:53,470
signs machine learning projects can be
2035
01:23:51,280 --> 01:23:55,930
in-house or outsourced I've seen both of
2036
01:23:53,470 --> 01:23:57,160
these models used successfully sometimes
2037
01:23:55,930 --> 01:23:59,950
if you outsource and machine learning
2038
01:23:57,160 --> 01:24:03,010
project you can have access much more
2039
01:23:59,950 --> 01:24:05,560
quickly to talent and get going faster
2040
01:24:03,010 --> 01:24:07,630
on the project it is nice if eventually
2041
01:24:05,560 --> 01:24:10,090
you build your own in-house AI team and
2042
01:24:07,630 --> 01:24:11,860
can also do these projects in-house you
2043
01:24:10,090 --> 01:24:14,980
hear more about this when we talk about
2044
01:24:11,860 --> 01:24:17,530
the AI translation playbook in greater
2045
01:24:14,980 --> 01:24:19,810
detail next week unlike machine learning
2046
01:24:17,530 --> 01:24:22,270
projects though data science projects
2047
01:24:19,810 --> 01:24:24,040
are more commonly done in-house they're
2048
01:24:22,270 --> 01:24:26,230
not impossible to outsource you can
2049
01:24:24,040 --> 01:24:28,510
sometimes outsource them but what I've
2050
01:24:26,230 --> 01:24:31,510
seen is that data science projects are
2051
01:24:28,510 --> 01:24:34,000
often so closely tied to your business
2052
01:24:31,510 --> 01:24:36,280
then it takes very deep day-to-day
2053
01:24:34,000 --> 01:24:38,680
knowledge about your business to do the
2054
01:24:36,280 --> 01:24:40,690
best data science projects and so just
2055
01:24:38,680 --> 01:24:42,820
as a percentage as a fraction
2056
01:24:40,690 --> 01:24:44,970
I see data science projects in house
2057
01:24:42,820 --> 01:24:48,880
more than machine learning projects
2058
01:24:44,970 --> 01:24:51,310
finally in every industry some things
2059
01:24:48,880 --> 01:24:54,370
will be industry standard and you should
2060
01:24:51,310 --> 01:24:57,100
avoid building those a common answer to
2061
01:24:54,370 --> 01:24:58,870
the build versus buy question was don't
2062
01:24:57,100 --> 01:25:00,520
the things they're going to be quite
2063
01:24:58,870 --> 01:25:02,320
specialized to you or completely
2064
01:25:00,520 --> 01:25:04,210
specialized to you or they'll allow you
2065
01:25:02,320 --> 01:25:07,240
to build a unique defensive advantage
2066
01:25:04,210 --> 01:25:09,610
but the things that will be industry
2067
01:25:07,240 --> 01:25:11,470
standard probably some other company
2068
01:25:09,610 --> 01:25:13,390
will build and it'll be more efficient
2069
01:25:11,470 --> 01:25:16,240
for you to just buy it rather than
2070
01:25:13,390 --> 01:25:19,060
building in hosts one of my team's had a
2071
01:25:16,240 --> 01:25:21,340
really poetic phrase which is don't
2072
01:25:19,060 --> 01:25:25,180
sprint in front of a train and what that
2073
01:25:21,340 --> 01:25:27,960
means is if this is a train running on
2074
01:25:25,180 --> 01:25:27,960
the railway tracks
2075
01:25:28,980 --> 01:25:34,750
and that's the nor chimney with the puff
2076
01:25:31,900 --> 01:25:37,210
of smoke what you don't want to do is to
2077
01:25:34,750 --> 01:25:39,489
be the person or the engineer trying to
2078
01:25:37,210 --> 01:25:41,619
sprint faster and faster ahead of the
2079
01:25:39,489 --> 01:25:44,770
Train the Train is the industry standard
2080
01:25:41,619 --> 01:25:46,989
solution and so if there's a company
2081
01:25:44,770 --> 01:25:48,880
maybe a Santa maybe a big company or
2082
01:25:46,989 --> 01:25:50,920
maybe an open-source effort that is
2083
01:25:48,880 --> 01:25:53,920
building an industry standard solution
2084
01:25:50,920 --> 01:25:55,989
then you may want to avoid trying to run
2085
01:25:53,920 --> 01:25:56,500
faster and faster to keep ahead of the
2086
01:25:55,989 --> 01:25:58,810
Train
2087
01:25:56,500 --> 01:26:01,420
because even though you could sprint
2088
01:25:58,810 --> 01:26:03,639
faster in the short term eventually the
2089
01:26:01,420 --> 01:26:05,469
train will catch up and you know crush
2090
01:26:03,639 --> 01:26:08,080
someone trying to sprint in front of a
2091
01:26:05,469 --> 01:26:10,360
train so when there's a massive force of
2092
01:26:08,080 --> 01:26:12,820
an industry standard solution that is
2093
01:26:10,360 --> 01:26:15,250
being built you might be better off just
2094
01:26:12,820 --> 01:26:17,530
embracing an industry standard or
2095
01:26:15,250 --> 01:26:19,719
embracing someone else's platform rather
2096
01:26:17,530 --> 01:26:21,159
than trying to do everything in-house
2097
01:26:19,719 --> 01:26:23,800
we are live in a world of limited
2098
01:26:21,159 --> 01:26:26,530
resources limited time limited data
2099
01:26:23,800 --> 01:26:28,989
limited and drain resources and so I
2100
01:26:26,530 --> 01:26:31,300
hope you can focus those resources on
2101
01:26:28,989 --> 01:26:33,040
the project so that most unique can make
2102
01:26:31,300 --> 01:26:35,380
the biggest difference to your company
2103
01:26:33,040 --> 01:26:38,020
through the process of technical
2104
01:26:35,380 --> 01:26:39,330
diligence as well as business diligence
2105
01:26:38,020 --> 01:26:41,920
I hope you can start to identify
2106
01:26:39,330 --> 01:26:44,679
projects that are potentially valuable
2107
01:26:41,920 --> 01:26:48,010
or that seem promising for your business
2108
01:26:44,679 --> 01:26:49,840
if the project is a big component maybe
2109
01:26:48,010 --> 01:26:51,969
I'll take many months to do it's not
2110
01:26:49,840 --> 01:26:53,949
unusual for me to spend even a few weeks
2111
01:26:51,969 --> 01:26:57,130
conducting this type of diligence before
2112
01:26:53,949 --> 01:26:59,800
committing to a project now say you've
2113
01:26:57,130 --> 01:27:01,719
found a few promising projects how do
2114
01:26:59,800 --> 01:27:03,580
you engage from an AI team how do you
2115
01:27:01,719 --> 01:27:05,199
work um in the I team to try to get
2116
01:27:03,580 --> 01:27:07,580
these projects done let's talk about
2117
01:27:05,199 --> 01:27:09,640
that in the next video
2118
01:27:07,580 --> 01:27:09,640
you
2119
01:27:11,250 --> 01:27:15,160
so you found
2120
01:27:13,120 --> 01:27:17,830
exciting project that you want to try to
2121
01:27:15,160 --> 01:27:20,500
excuse on how do you work of an AI team
2122
01:27:17,830 --> 01:27:22,540
on this project in this video you
2123
01:27:20,500 --> 01:27:25,060
learned how a I teams think about data
2124
01:27:22,540 --> 01:27:26,980
and therefore how you can interact with
2125
01:27:25,060 --> 01:27:30,190
AI teams to help them succeed on a
2126
01:27:26,980 --> 01:27:32,110
project now there is one caveat which is
2127
01:27:30,190 --> 01:27:33,910
what if you have a cool idea but you
2128
01:27:32,110 --> 01:27:35,410
don't have access to an AI team you
2129
01:27:33,910 --> 01:27:36,360
don't have any access to any AI
2130
01:27:35,410 --> 01:27:38,560
engineers
2131
01:27:36,360 --> 01:27:41,230
fortunately in today's world if either
2132
01:27:38,560 --> 01:27:43,150
you yourself or you can encourage some
2133
01:27:41,230 --> 01:27:45,580
of your injuring your friends to take an
2134
01:27:43,150 --> 01:27:47,680
online course or two on machine learning
2135
01:27:45,580 --> 01:27:49,750
or deep learning that often will give
2136
01:27:47,680 --> 01:27:51,190
them enough knowledge to get going and
2137
01:27:49,750 --> 01:27:52,840
make a start of an attempt make a
2138
01:27:51,190 --> 01:27:55,600
reasonable attempt on these types of
2139
01:27:52,840 --> 01:27:58,990
projects so let's talk about how you can
2140
01:27:55,600 --> 01:28:01,630
work with an AI team first it really
2141
01:27:58,990 --> 01:28:04,300
helps your AI team if you can specify an
2142
01:28:01,630 --> 01:28:06,730
acceptance criteria for the project I've
2143
01:28:04,300 --> 01:28:09,160
done a lot of work in automated visual
2144
01:28:06,730 --> 01:28:11,460
inspection so I'm going to use that as a
2145
01:28:09,160 --> 01:28:14,740
running example in these few slides
2146
01:28:11,460 --> 01:28:17,290
let's say your goal is to detect defects
2147
01:28:14,740 --> 01:28:19,810
in coffee mugs with at least 95 percent
2148
01:28:17,290 --> 01:28:23,680
accuracy so that can be your acceptance
2149
01:28:19,810 --> 01:28:27,040
criteria for this project but 95 percent
2150
01:28:23,680 --> 01:28:28,720
accuracy how do you measure accuracy one
2151
01:28:27,040 --> 01:28:31,960
of the things that the AI team would
2152
01:28:28,720 --> 01:28:35,290
need is a data set on which to measure
2153
01:28:31,960 --> 01:28:37,810
their accuracy so data set is just a set
2154
01:28:35,290 --> 01:28:40,990
of pictures like these together with the
2155
01:28:37,810 --> 01:28:43,060
labels with the design output be that
2156
01:28:40,990 --> 01:28:46,720
the first two coffee mugs are okay and
2157
01:28:43,060 --> 01:28:48,760
the third one is defective so as part of
2158
01:28:46,720 --> 01:28:51,190
your specification for the acceptance
2159
01:28:48,760 --> 01:28:53,860
criteria you should make sure that the
2160
01:28:51,190 --> 01:28:55,960
AI team has a data set on which to
2161
01:28:53,860 --> 01:28:58,300
measure the performance so that they can
2162
01:28:55,960 --> 01:29:01,780
know if they've achieved 95 percent
2163
01:28:58,300 --> 01:29:06,580
accuracy the formal term for this data
2164
01:29:01,780 --> 01:29:09,580
set is called a test set and the test
2165
01:29:06,580 --> 01:29:11,860
set may not need to be too big maybe a
2166
01:29:09,580 --> 01:29:14,320
thousand pictures will be just fine for
2167
01:29:11,860 --> 01:29:16,270
this example but if you consulted that
2168
01:29:14,320 --> 01:29:18,880
AI expert they can give you a better
2169
01:29:16,270 --> 01:29:20,800
sense of how big the test set needs to
2170
01:29:18,880 --> 01:29:23,500
be for them to be able to evaluate
2171
01:29:20,800 --> 01:29:26,680
whether or not they're getting to 95%
2172
01:29:23,500 --> 01:29:29,470
accuracy one novel part of a
2173
01:29:26,680 --> 01:29:32,290
systems is that the performance is
2174
01:29:29,470 --> 01:29:34,900
usually specified in a statistical way
2175
01:29:32,290 --> 01:29:37,420
so rather than a free friend AI system
2176
01:29:34,900 --> 01:29:39,550
that just does something perfectly you
2177
01:29:37,420 --> 01:29:41,800
see very often that we want any AI
2178
01:29:39,550 --> 01:29:44,170
system that performs at a certain
2179
01:29:41,800 --> 01:29:46,840
percentage accuracy like this example
2180
01:29:44,170 --> 01:29:49,090
here so when specifying your acceptance
2181
01:29:46,840 --> 01:29:50,830
criteria think of whether your
2182
01:29:49,090 --> 01:29:53,650
acceptance criteria needs to be
2183
01:29:50,830 --> 01:29:56,740
specified in a statistical way where you
2184
01:29:53,650 --> 01:29:58,600
specify on average our does or what
2185
01:29:56,740 --> 01:30:01,120
percent of time it has to get the right
2186
01:29:58,600 --> 01:30:04,570
answer let's dive more deeply into the
2187
01:30:01,120 --> 01:30:07,150
concept of a test set this is how AI
2188
01:30:04,570 --> 01:30:10,000
teams think about data AI teams group
2189
01:30:07,150 --> 01:30:12,370
data into two main data says the first
2190
01:30:10,000 --> 01:30:14,290
called the training set and the second
2191
01:30:12,370 --> 01:30:16,690
called the test set which we've already
2192
01:30:14,290 --> 01:30:18,580
talked a bit about the training set is
2193
01:30:16,690 --> 01:30:20,830
just a set of pictures together with
2194
01:30:18,580 --> 01:30:23,230
labels showing whether each of these
2195
01:30:20,830 --> 01:30:26,230
pictures is of a coffee mug that is okay
2196
01:30:23,230 --> 01:30:28,690
or defective so the training set gives
2197
01:30:26,230 --> 01:30:30,670
examples of both the input a the
2198
01:30:28,690 --> 01:30:33,670
pictures of the coffee mouse as well as
2199
01:30:30,670 --> 01:30:36,910
the desired output B whether it's okay
2200
01:30:33,670 --> 01:30:40,420
or defective and so given this training
2201
01:30:36,910 --> 01:30:43,900
set what a machine learning algorithm
2202
01:30:40,420 --> 01:30:47,350
will do is learn in other words compute
2203
01:30:43,900 --> 01:30:49,990
or figure out some mapping from A to B
2204
01:30:47,350 --> 01:30:52,480
so that you now have a piece of software
2205
01:30:49,990 --> 01:30:54,370
they can take as input the input a and
2206
01:30:52,480 --> 01:30:56,890
try to figure out what is the
2207
01:30:54,370 --> 01:30:58,840
appropriate output B so the training set
2208
01:30:56,890 --> 01:31:02,020
is the input to the machine learning
2209
01:30:58,840 --> 01:31:04,810
software that lets it figure out what is
2210
01:31:02,020 --> 01:31:07,210
this a to b mapping the second data said
2211
01:31:04,810 --> 01:31:09,610
that an AI team will use is the test set
2212
01:31:07,210 --> 01:31:11,740
and as you've seen this is just another
2213
01:31:09,610 --> 01:31:13,960
set of images that's different from the
2214
01:31:11,740 --> 01:31:16,390
training set also what the labels
2215
01:31:13,960 --> 01:31:18,430
provided the way an AI team will
2216
01:31:16,390 --> 01:31:21,730
evaluate their learning algorithms
2217
01:31:18,430 --> 01:31:24,880
performance is to give the images in the
2218
01:31:21,730 --> 01:31:27,160
test set to the AI software and see what
2219
01:31:24,880 --> 01:31:29,920
the AI software outputs for example if
2220
01:31:27,160 --> 01:31:32,860
on these three tests set images the AI
2221
01:31:29,920 --> 01:31:36,010
software outputs okay for this okay for
2222
01:31:32,860 --> 01:31:38,170
this and also okay for this then we will
2223
01:31:36,010 --> 01:31:39,960
say that they got two out of three
2224
01:31:38,170 --> 01:31:43,830
examples right
2225
01:31:39,960 --> 01:31:46,890
so that's a 66.7%
2226
01:31:43,830 --> 01:31:48,960
accuracy in this figure the training set
2227
01:31:46,890 --> 01:31:51,480
and test sets are both only three
2228
01:31:48,960 --> 01:31:53,550
pictures in practice both of these data
2229
01:31:51,480 --> 01:31:55,680
sets would be much bigger of course and
2230
01:31:53,550 --> 01:31:57,630
you find it for most problems the
2231
01:31:55,680 --> 01:32:00,300
training set is much much much bigger
2232
01:31:57,630 --> 01:32:02,460
than the test set but you can talk to AI
2233
01:32:00,300 --> 01:32:04,280
engineers to find out how much data they
2234
01:32:02,460 --> 01:32:07,620
need for a given problem
2235
01:32:04,280 --> 01:32:09,900
finally for technical reasons some AI
2236
01:32:07,620 --> 01:32:12,570
teams will need not just one but two
2237
01:32:09,900 --> 01:32:14,969
different test sets if you hear AI teams
2238
01:32:12,570 --> 01:32:17,340
talk about development or dev or
2239
01:32:14,969 --> 01:32:19,920
validation sets that's the second test
2240
01:32:17,340 --> 01:32:21,690
set that they're using the reasons why
2241
01:32:19,920 --> 01:32:23,580
they need to test says is quite
2242
01:32:21,690 --> 01:32:26,100
technical and beyond the scope of this
2243
01:32:23,580 --> 01:32:27,750
course but if an AI team asks you for
2244
01:32:26,100 --> 01:32:29,430
two different test sets is quite
2245
01:32:27,750 --> 01:32:31,830
reasonable to try to provide that to
2246
01:32:29,430 --> 01:32:34,230
them before wrapping up this video one
2247
01:32:31,830 --> 01:32:36,300
pitfall I want to urge you to avoid is
2248
01:32:34,230 --> 01:32:38,910
expecting a hundred percent accuracy
2249
01:32:36,300 --> 01:32:41,670
from your AI software here's what I mean
2250
01:32:38,910 --> 01:32:43,410
let's say this is your test set which
2251
01:32:41,670 --> 01:32:46,770
you've already seen on the last slide
2252
01:32:43,410 --> 01:32:48,750
but let me add a few more examples to
2253
01:32:46,770 --> 01:32:50,790
this test set here are some of the
2254
01:32:48,750 --> 01:32:52,830
reasons it may not be possible for a
2255
01:32:50,790 --> 01:32:55,590
piece of AI software to be a hundred
2256
01:32:52,830 --> 01:32:57,630
percent accurate first machine learning
2257
01:32:55,590 --> 01:33:00,030
technology today despite being very
2258
01:32:57,630 --> 01:33:02,370
powerful slow has limitations and they
2259
01:33:00,030 --> 01:33:04,050
just can't do everything and so you may
2260
01:33:02,370 --> 01:33:06,270
be working on a problem that is just
2261
01:33:04,050 --> 01:33:09,570
very difficult even for today's machine
2262
01:33:06,270 --> 01:33:11,840
learning technology second insufficient
2263
01:33:09,570 --> 01:33:14,010
data if you don't have enough data
2264
01:33:11,840 --> 01:33:16,590
specifically if you don't have enough
2265
01:33:14,010 --> 01:33:18,900
training data for the AI software to
2266
01:33:16,590 --> 01:33:22,440
learn from it may be very difficult to
2267
01:33:18,900 --> 01:33:25,440
get a very high level of accuracy third
2268
01:33:22,440 --> 01:33:26,340
data is messy and sometimes data can be
2269
01:33:25,440 --> 01:33:28,680
mislabeled
2270
01:33:26,340 --> 01:33:33,210
for example this green coffee mug here
2271
01:33:28,680 --> 01:33:35,190
looks perfectly okay to me so the label
2272
01:33:33,210 --> 01:33:38,340
of it being a defect looks like an
2273
01:33:35,190 --> 01:33:41,670
incorrect label and that would hurt the
2274
01:33:38,340 --> 01:33:44,850
performance of your AI software and data
2275
01:33:41,670 --> 01:33:47,130
can also be ambiguous for example it
2276
01:33:44,850 --> 01:33:49,200
looks like this coffee mug has a small
2277
01:33:47,130 --> 01:33:51,750
scratch over there and it's a pretty
2278
01:33:49,200 --> 01:33:52,450
small scratch so maybe we will think of
2279
01:33:51,750 --> 01:33:55,270
this though
2280
01:33:52,450 --> 01:33:58,570
hey that maybe this should actually have
2281
01:33:55,270 --> 01:34:00,940
been a defect or maybe even different
2282
01:33:58,570 --> 01:34:03,520
experts won't agree if this book of
2283
01:34:00,940 --> 01:34:05,410
coffee mug is okay I should pass the
2284
01:34:03,520 --> 01:34:08,740
inspection step some of these problems
2285
01:34:05,410 --> 01:34:10,990
can be ameliorated for example if you
2286
01:34:08,740 --> 01:34:13,390
don't have enough data maybe you can try
2287
01:34:10,990 --> 01:34:16,480
to collect more data and more data more
2288
01:34:13,390 --> 01:34:19,330
often help or you can also try to clean
2289
01:34:16,480 --> 01:34:21,400
up mislabeled data or try to get your
2290
01:34:19,330 --> 01:34:23,470
factories expensed come to better
2291
01:34:21,400 --> 01:34:25,390
agreement about these ambiguous labels
2292
01:34:23,470 --> 01:34:28,870
so there are ways to try to make these
2293
01:34:25,390 --> 01:34:31,870
things better but a lot of AI systems
2294
01:34:28,870 --> 01:34:34,720
are incredibly valuable even without
2295
01:34:31,870 --> 01:34:36,610
achieving a hundred percent accuracy so
2296
01:34:34,720 --> 01:34:38,920
I would urge you to discuss with your AI
2297
01:34:36,610 --> 01:34:41,380
engineers what is a reasonable level of
2298
01:34:38,920 --> 01:34:43,450
accuracy to try to accomplish and then
2299
01:34:41,380 --> 01:34:46,480
try to find something that passes both
2300
01:34:43,450 --> 01:34:48,820
technical diligence as well as business
2301
01:34:46,480 --> 01:34:51,670
diligence without necessarily needing a
2302
01:34:48,820 --> 01:34:53,470
hundred percent accuracy congratulations
2303
01:34:51,670 --> 01:34:55,900
on finishing all the videos for this
2304
01:34:53,470 --> 01:34:58,360
week you now know what it feels like and
2305
01:34:55,900 --> 01:34:59,920
what it takes to build an AI project and
2306
01:34:58,360 --> 01:35:02,770
I hope you start brainstorming and
2307
01:34:59,920 --> 01:35:05,200
exploring some ideas there is one more
2308
01:35:02,770 --> 01:35:07,660
optional video describing some of the
2309
01:35:05,200 --> 01:35:09,910
technical tools that AI teams use they
2310
01:35:07,660 --> 01:35:11,830
can watch if you wish but either way I
2311
01:35:09,910 --> 01:35:14,650
look forward to seeing you next week
2312
01:35:11,830 --> 01:35:16,810
where you learn how a AI projects fit in
2313
01:35:14,650 --> 01:35:20,160
the context of a bigger company look
2314
01:35:16,810 --> 01:35:20,160
forward to seeing you next week
2315
01:35:22,679 --> 01:35:27,510
when you work with AI teams you may hear
2316
01:35:25,260 --> 01:35:30,570
them refer to the tools that they're
2317
01:35:27,510 --> 01:35:32,400
using to build these AI systems in this
2318
01:35:30,570 --> 01:35:34,229
video I want to share of you some
2319
01:35:32,400 --> 01:35:36,959
details and names of the most commonly
2320
01:35:34,229 --> 01:35:39,150
used AI tools so that you people better
2321
01:35:36,959 --> 01:35:41,670
understand what these AI engineers are
2322
01:35:39,150 --> 01:35:44,249
doing we're fortunate that the AI world
2323
01:35:41,670 --> 01:35:47,489
today is very open and many teams will
2324
01:35:44,249 --> 01:35:50,249
openly share ideas of each other there
2325
01:35:47,489 --> 01:35:53,189
are great machine learning open source
2326
01:35:50,249 --> 01:35:55,349
frameworks that many teams are used to
2327
01:35:53,189 --> 01:35:58,110
build their systems so if you hear of
2328
01:35:55,349 --> 01:36:00,209
any of these 10sec philip I don't care
2329
01:35:58,110 --> 01:36:02,880
is MX nets the antique a cafe paddle
2330
01:36:00,209 --> 01:36:04,769
paddle so I can learn our or Wecker all
2331
01:36:02,880 --> 01:36:07,679
of these are open source machine
2332
01:36:04,769 --> 01:36:09,900
learning frameworks that help AI teams
2333
01:36:07,679 --> 01:36:12,599
be much more efficient in terms of
2334
01:36:09,900 --> 01:36:14,789
writing software along of AI technology
2335
01:36:12,599 --> 01:36:17,400
breakthroughs are also published freely
2336
01:36:14,789 --> 01:36:20,219
on the internet on this website called
2337
01:36:17,400 --> 01:36:23,340
archive it's felt like this I hope that
2338
01:36:20,219 --> 01:36:25,289
other academic communities also freely
2339
01:36:23,340 --> 01:36:27,389
share their research since I've seen
2340
01:36:25,289 --> 01:36:29,329
firsthand how much does accelerates
2341
01:36:27,389 --> 01:36:32,909
progress in the whole field of AI
2342
01:36:29,329 --> 01:36:35,249
finally many teams will also share their
2343
01:36:32,909 --> 01:36:37,320
code freely on the internet most
2344
01:36:35,249 --> 01:36:40,679
commonly on the website called github
2345
01:36:37,320 --> 01:36:43,559
this has become the de facto repository
2346
01:36:40,679 --> 01:36:46,340
for open-source software in AI and in
2347
01:36:43,559 --> 01:36:48,389
other sectors in AI and by using
2348
01:36:46,340 --> 01:36:51,119
appropriately licensed open-source
2349
01:36:48,389 --> 01:36:52,860
software many teams can get going much
2350
01:36:51,119 --> 01:36:55,469
faster than if they had to build
2351
01:36:52,860 --> 01:36:59,760
everything from scratch so for example
2352
01:36:55,469 --> 01:37:06,719
if I search online for face recognition
2353
01:36:59,760 --> 01:37:10,130
software on github you might find a web
2354
01:37:06,719 --> 01:37:13,439
page like this and if you scroll down
2355
01:37:10,130 --> 01:37:16,469
this actually has a pretty good very
2356
01:37:13,439 --> 01:37:18,300
readable description of software that is
2357
01:37:16,469 --> 01:37:20,969
made available on this website for
2358
01:37:18,300 --> 01:37:24,570
recognizing people's faces and even
2359
01:37:20,969 --> 01:37:27,869
finding parts of people's faces there's
2360
01:37:24,570 --> 01:37:29,909
just a ton of software that is freely
2361
01:37:27,869 --> 01:37:31,920
downloadable for doing all sorts of
2362
01:37:29,909 --> 01:37:33,719
things on the internet and just double
2363
01:37:31,920 --> 01:37:36,330
check the license or AI team would
2364
01:37:33,719 --> 01:37:38,790
double check the license before using it
2365
01:37:36,330 --> 01:37:41,610
of course but a lot of the software is
2366
01:37:38,790 --> 01:37:43,710
open source or otherwise very pretty
2367
01:37:41,610 --> 01:37:46,890
mystically license for anyone to use
2368
01:37:43,710 --> 01:37:49,710
although github is a technical website
2369
01:37:46,890 --> 01:37:51,420
built for engineers if you want you
2370
01:37:49,710 --> 01:37:53,670
should feel free to play around github
2371
01:37:51,420 --> 01:37:56,070
and see what are the types of AI
2372
01:37:53,670 --> 01:37:58,560
software people have released online as
2373
01:37:56,070 --> 01:38:01,440
well in addition to these open source
2374
01:37:58,560 --> 01:38:03,890
technical tools you often also hear AI
2375
01:38:01,440 --> 01:38:07,140
engineers talk about CPUs and GPUs
2376
01:38:03,890 --> 01:38:10,170
here's what these terms mean a CPU is
2377
01:38:07,140 --> 01:38:12,900
the computer processor in your computer
2378
01:38:10,170 --> 01:38:15,870
whether is your desktop your laptop or a
2379
01:38:12,900 --> 01:38:18,410
computer server off in the cloud CPU
2380
01:38:15,870 --> 01:38:22,590
stands for a central processing unit and
2381
01:38:18,410 --> 01:38:24,810
CPUs are made by Intel and AMD and a few
2382
01:38:22,590 --> 01:38:28,980
other companies this does a lot of the
2383
01:38:24,810 --> 01:38:31,310
computation in your computer GPU stands
2384
01:38:28,980 --> 01:38:34,800
for graphics processing unit
2385
01:38:31,310 --> 01:38:38,610
historically the GPU was made to process
2386
01:38:34,800 --> 01:38:41,520
pictures so if you play a video gaem is
2387
01:38:38,610 --> 01:38:44,160
probably a GPU that is drawing the fancy
2388
01:38:41,520 --> 01:38:46,110
graphics but what we found several years
2389
01:38:44,160 --> 01:38:48,350
ago was that the hardware there was
2390
01:38:46,110 --> 01:38:51,270
originally built for processing graphics
2391
01:38:48,350 --> 01:38:54,300
turns out to be very very powerful for
2392
01:38:51,270 --> 01:38:57,060
building very large new networks or very
2393
01:38:54,300 --> 01:38:59,280
large deep learning algorithms given the
2394
01:38:57,060 --> 01:39:02,580
need to build very large deep learning
2395
01:38:59,280 --> 01:39:04,920
or very large neural network systems the
2396
01:39:02,580 --> 01:39:07,770
AI community has had this insatiable
2397
01:39:04,920 --> 01:39:09,810
hunger for more and more computational
2398
01:39:07,770 --> 01:39:12,330
power to train bigger and bigger neural
2399
01:39:09,810 --> 01:39:14,460
networks and GPUs have proved to be a
2400
01:39:12,330 --> 01:39:16,800
fantastic fit to this type of
2401
01:39:14,460 --> 01:39:19,650
computation that we need to have done to
2402
01:39:16,800 --> 01:39:22,320
train very large neural networks so
2403
01:39:19,650 --> 01:39:25,800
that's why GPUs are playing a big role
2404
01:39:22,320 --> 01:39:27,600
in the rise of deep learning and then
2405
01:39:25,800 --> 01:39:30,510
there is a company that's been selling
2406
01:39:27,600 --> 01:39:33,450
many GPUs but other companies including
2407
01:39:30,510 --> 01:39:35,750
Qualcomm as well as Google making his
2408
01:39:33,450 --> 01:39:38,610
own TP use are increasingly making
2409
01:39:35,750 --> 01:39:41,310
specialized hardware for powering these
2410
01:39:38,610 --> 01:39:43,190
very large neural networks finally you
2411
01:39:41,310 --> 01:39:46,230
might hear about cloud versus
2412
01:39:43,190 --> 01:39:48,970
on-premises or for short on the Prem
2413
01:39:46,230 --> 01:39:51,850
deployments cloud deployments refer to
2414
01:39:48,970 --> 01:39:55,390
if you rent compute service such as from
2415
01:39:51,850 --> 01:39:58,630
Amazon's AWS or Microsoft Azure or
2416
01:39:55,390 --> 01:40:01,180
Google's GCP in order to use someone
2417
01:39:58,630 --> 01:40:03,820
else's service to do your computation
2418
01:40:01,180 --> 01:40:05,680
whereas an on-prem deployment means
2419
01:40:03,820 --> 01:40:08,110
buying your own compute service and
2420
01:40:05,680 --> 01:40:11,020
running the service locally in your own
2421
01:40:08,110 --> 01:40:13,180
company a detailed exploration of the
2422
01:40:11,020 --> 01:40:15,550
pros and cons of these two options is
2423
01:40:13,180 --> 01:40:17,860
beyond the scope of this video a lot of
2424
01:40:15,550 --> 01:40:19,840
the world is moving to cloud deployments
2425
01:40:17,860 --> 01:40:22,060
but the research online do you find many
2426
01:40:19,840 --> 01:40:24,520
articles talking about the pros and cons
2427
01:40:22,060 --> 01:40:27,250
of cloud versus on-prem deployments
2428
01:40:24,520 --> 01:40:30,700
there is one last term you might hear
2429
01:40:27,250 --> 01:40:32,920
about which is edge deployments if you
2430
01:40:30,700 --> 01:40:35,050
are building a self-driving car there's
2431
01:40:32,920 --> 01:40:37,720
not enough time to send data from a
2432
01:40:35,050 --> 01:40:39,430
self-driving car to a cloud server to
2433
01:40:37,720 --> 01:40:40,990
decide if you can stop the car or not
2434
01:40:39,430 --> 01:40:43,420
and then send that message back to the
2435
01:40:40,990 --> 01:40:46,030
self-driving car so the computation has
2436
01:40:43,420 --> 01:40:48,880
to happen usually in the computer right
2437
01:40:46,030 --> 01:40:51,640
there inside the car that's called an H
2438
01:40:48,880 --> 01:40:54,520
deployment where you put a processor
2439
01:40:51,640 --> 01:40:56,380
right where the data is collected so
2440
01:40:54,520 --> 01:40:57,610
that you can process the data and make a
2441
01:40:56,380 --> 01:41:00,250
decision very quickly
2442
01:40:57,610 --> 01:41:02,770
without needing to transmit the data
2443
01:41:00,250 --> 01:41:05,410
over the internet to be processed
2444
01:41:02,770 --> 01:41:07,180
somewhere else if you look at some of
2445
01:41:05,410 --> 01:41:10,930
the small speakers in your home as well
2446
01:41:07,180 --> 01:41:12,820
this too is an H deployment where some
2447
01:41:10,930 --> 01:41:15,760
not all but some of the speech
2448
01:41:12,820 --> 01:41:18,160
recognition toss is done by a processor
2449
01:41:15,760 --> 01:41:20,020
that is built-in right there into the
2450
01:41:18,160 --> 01:41:23,410
small speaker that is inside your home
2451
01:41:20,020 --> 01:41:26,050
the main advantage of a deployment is it
2452
01:41:23,410 --> 01:41:28,690
can increase response time of the system
2453
01:41:26,050 --> 01:41:31,000
and also reduce the amount of data you
2454
01:41:28,690 --> 01:41:33,430
need to send over the network but there
2455
01:41:31,000 --> 01:41:35,520
are many pros and cons as well about a
2456
01:41:33,430 --> 01:41:38,230
tree versus cloud versus on-prem
2457
01:41:35,520 --> 01:41:40,750
deployments that you can also search
2458
01:41:38,230 --> 01:41:42,850
online to read more about thanks for
2459
01:41:40,750 --> 01:41:44,770
finishing this optional video on the
2460
01:41:42,850 --> 01:41:46,870
technical tools that AI engineers use
2461
01:41:44,770 --> 01:41:48,400
hopefully when you hear them refer to
2462
01:41:46,870 --> 01:41:50,560
some of these tools you start to have a
2463
01:41:48,400 --> 01:41:53,910
better sense of what they mean I look
2464
01:41:50,560 --> 01:41:53,910
forward to seeing you next week190191
Can't find what you're looking for?
Get subtitles in any language from opensubtitles.com, and translate them here.