1
00:00:01,130 --> 00:00:20,270
So said. I am deeply humbled to be here tonight, and I'd like to, if I do well.

2
00:00:21,200 --> 00:00:26,690
Which is a big if. Dedicate this lecture tonight to my.

3
00:00:27,760 --> 00:00:30,910
High school teacher. And here is why.

4
00:00:31,740 --> 00:00:45,150
Many years ago, more than three decades ago, when I was in high school, I was selected to prepare for and then to compete in the Mathematics Olympics.

5
00:00:47,080 --> 00:00:53,110
Because they thought people thought in Austria that I was smart enough to be a good contestant.

6
00:00:56,140 --> 00:01:05,420
And to my embarrassment, they didn't do very well. So my high school teacher, Hans Kraus, came to me and he said,

7
00:01:07,100 --> 00:01:14,929
You're just not smart enough for a real good mathematician, but maybe you can be an applied mathematician.

8
00:01:14,930 --> 00:01:21,410
So how about the how about physics Olympics?

9
00:01:25,940 --> 00:01:30,090
So I did. And I did very well. And.

10
00:01:33,000 --> 00:01:43,590
With that, I shall start and dedicate this lecture to Hans Kraus, who gave me the guts to speak to an audience of mathematicians today.

11
00:01:45,930 --> 00:01:49,230
I'm starting to talk about big data with a case study,

12
00:01:49,890 --> 00:01:53,970
a particular case that has to do with something that you're all familiar with, and that is the flu.

13
00:01:54,780 --> 00:02:00,090
Every year, tens of thousands of people around the world die of the seasonal flu.

14
00:02:00,840 --> 00:02:06,720
But in 2009, a new flu virus was detected, the H1N1 virus.

15
00:02:07,260 --> 00:02:10,110
And at that stage, there was no vaccine available.

16
00:02:10,890 --> 00:02:18,480
So the best that public health authorities around the world could do was to try to limit the spread of the virus.

17
00:02:18,480 --> 00:02:22,320
But for that, they needed to know where the virus was.

18
00:02:23,830 --> 00:02:31,120
In the United States. This role is taken on by the Centres for Disease Control in Atlanta and they have doctors,

19
00:02:31,120 --> 00:02:37,300
general practitioners from around the country to tell them each and every H1N1 case that they have.

20
00:02:39,320 --> 00:02:48,440
And based on that and some hard labour over days and days and days, they're able to tell the policymakers.

21
00:02:49,950 --> 00:02:53,790
Were the flu is at any given point in time.

22
00:02:54,700 --> 00:02:58,050
Two weeks later. Which is an eternity.

23
00:02:58,350 --> 00:03:04,390
If you have a pandemic underway. Just around the same time.

24
00:03:05,280 --> 00:03:14,520
Engineers at a little Start-Up Company called Google in Mountain View had a very different idea of how to predict the spread of the flu.

25
00:03:14,790 --> 00:03:21,460
They said, we'd like to predict the spread of the flu by just using Google searches.

26
00:03:21,480 --> 00:03:33,400
That is search requests sent to Google. Google receives about 5 billion search requests every single day and has stored and saved all of them,

27
00:03:33,400 --> 00:03:37,440
including information where they came from and so forth over the last 15 years.

28
00:03:37,810 --> 00:03:44,440
So the idea was to take the last five years or so of Google search requests that it received and to take

29
00:03:44,440 --> 00:03:50,560
official figures from the Centres for Disease Control and to see whether there would be a correlation,

30
00:03:51,460 --> 00:04:04,370
some kind of a connection. And in that process, Google's engineers tested 50 million different search terms for 150 million mathematical models.

31
00:04:04,940 --> 00:04:11,300
And then they found one that provided a pretty darn good prediction.

32
00:04:12,970 --> 00:04:18,010
Here you see the official CDC data and the Google Trends predictions.

33
00:04:18,960 --> 00:04:28,980
But. Importantly were the CDC prediction was always about 10 to 14 days late.

34
00:04:29,730 --> 00:04:33,720
Google could do it almost in real time. And.

35
00:04:34,740 --> 00:04:40,350
The idea behind this, the direction that this points towards is big data.

36
00:04:40,680 --> 00:04:47,850
Now, big data is not just about public health. We find it everywhere, including in finance.

37
00:04:48,650 --> 00:05:00,020
In financial services. The vice chancellor has already alluded to, I think, in his speech, the importance of this matter,

38
00:05:00,260 --> 00:05:08,840
especially in light of the Great Recession of 2008, when we look at September of 2008.

39
00:05:11,150 --> 00:05:16,610
And when we now read the protocols. Of the Open Market Committee.

40
00:05:17,650 --> 00:05:25,720
Of the Federal Reserve. We find that they were clueless about what was going on because the data was not there.

41
00:05:26,620 --> 00:05:34,570
In fact, the deflationary tack after the Lehman Brothers default, after the Lehman Brother bankruptcy,

42
00:05:35,110 --> 00:05:44,350
was only felt in the CPI and the Consumer Price Index many weeks later because it takes many weeks to do the Consumer Price Index.

43
00:05:46,680 --> 00:05:52,590
There's a small Start-Up Company out of a research project at the Massachusetts Institute of Technology.

44
00:05:52,980 --> 00:06:02,820
Start-Up Company now is called Price Stats. And what they do is they go out on the Internet and every single day, every single hour, they suck down.

45
00:06:03,870 --> 00:06:10,770
Over a billion price points of consumer goods from hundreds of thousands of different offerings.

46
00:06:12,030 --> 00:06:18,870
And they do this in order to get very early indications of inflationary or deflationary trends.

47
00:06:20,350 --> 00:06:24,720
And so they were able to. See.

48
00:06:25,870 --> 00:06:32,950
The deflationary impact of Lehman Brothers and what came afterwards much earlier, if in fact.

49
00:06:33,890 --> 00:06:43,880
Policymakers had known if in fact, Federal Reserve policymakers had known perhaps some of the policy decisions would have been made differently.

50
00:06:44,660 --> 00:06:52,460
It is big data. The idea that we can gain insights from a large number of data points that we couldn't otherwise.

51
00:06:52,520 --> 00:06:58,250
Now, all three human history. We human beings have tried to make sense of the world around us.

52
00:06:59,950 --> 00:07:08,660
And we did that mostly by observing it. And in order to do that, we needed to collect data.

53
00:07:10,550 --> 00:07:21,680
But for all of human history, until extremely recently, the collection of data was hard, difficult, time consuming, expensive.

54
00:07:22,160 --> 00:07:30,860
And so in this world of small data, we collected as little data as absolutely necessary in order to answer the question that we had.

55
00:07:31,160 --> 00:07:41,480
We defined, designed, created the processes, the institutions, the structures that we used in order to make sense of the data.

56
00:07:42,050 --> 00:07:51,910
Knowing that we had very little of it. And so that we needed to squeeze every last drop of meaning out of them.

57
00:07:53,060 --> 00:07:58,580
What if that changes? What if we could envision a world in which there is.

58
00:08:02,800 --> 00:08:07,510
A lot of data available. Then we would have to rethink.

59
00:08:08,800 --> 00:08:13,390
Perhaps the processes and the institutions and the structures that we use to make sense of the world,

60
00:08:13,660 --> 00:08:19,270
much like this gentleman who is speaking out of the old world view and into a new one.

61
00:08:22,520 --> 00:08:29,210
It all began perhaps some 15, 20 years ago in earnest in the natural sciences.

62
00:08:29,660 --> 00:08:38,150
Consider astronomy. When the Sloan Digital Sky Survey came online in the year 2000, it collected.

63
00:08:39,400 --> 00:08:47,710
More data in its first few weeks of operation than had been amassed in the entire history of astronomy.

64
00:08:49,660 --> 00:08:58,030
Since then, since the year 2000, it has amassed over 200 terabytes of astronomy data, 200 terabytes of astronomy.

65
00:08:59,240 --> 00:09:07,880
But a successor telescope, you see a rendering that is going to go on stream in the year 2016 is going to collect that much data.

66
00:09:08,800 --> 00:09:14,070
Every five days. Or take genetics.

67
00:09:14,970 --> 00:09:20,190
In April of 2003, the world celebrated a tremendous achievement.

68
00:09:20,760 --> 00:09:25,440
That is, after ten years, $1 billion and a global effort.

69
00:09:25,740 --> 00:09:32,340
We, the world, were able to have a full sequence of one person's DNA.

70
00:09:32,640 --> 00:09:39,400
3 billion base pairs. Fast forward to today and sequencing my DNA.

71
00:09:39,940 --> 00:09:47,230
Not just one person's DNA. My DNA costs about less than $1,000 and takes 2 to 3 days in one lab.

72
00:09:48,760 --> 00:10:00,280
And that generates another 3 billion base pairs of data Internet companies to a drowning in data, 500 million tweets with Twitter every single day.

73
00:10:00,940 --> 00:10:05,530
800 million YouTube users upload an hour of video every single second.

74
00:10:05,680 --> 00:10:09,580
Even if you stop sleeping, you could never watch that amount of video.

75
00:10:11,110 --> 00:10:15,760
And on Facebook, 10 million photos are uploaded every single hour.

76
00:10:17,170 --> 00:10:21,430
Google processes dozens of petabytes of data every single day.

77
00:10:22,360 --> 00:10:25,810
Dozens of petabytes. Petabyte.

78
00:10:25,990 --> 00:10:30,040
Petabyte. Petabyte. Petabyte. What did you have for lunch?

79
00:10:30,430 --> 00:10:36,910
A couple of petabytes. How much is a freaking petabyte of data?

80
00:10:38,430 --> 00:10:50,489
If you take all of the characters in a book and all of the books and magazines and all of the other holdings of the largest library in the world,

81
00:10:50,490 --> 00:10:55,500
the Library of Congress together, and then multiply it by 100.

82
00:10:56,040 --> 00:11:03,210
That's about a petabyte. If we look at the growth of data in the world,

83
00:11:03,450 --> 00:11:14,610
the best guesstimate that we have over time tells us that the amount of data in the world from 1987 to 2007 grew 100 fold.

84
00:11:17,560 --> 00:11:21,590
Now. That's quite amazing, isn't it?

85
00:11:21,980 --> 00:11:29,390
100 times increase if Elizabeth Eisenstein maintains we go back in human history to find.

86
00:11:30,410 --> 00:11:41,350
A time when data are increased as much. We have to go back to 1450 to a rough 1506 in these 56 years of the Gutenberg revolution.

87
00:11:43,260 --> 00:11:50,460
The amount of data in the world doubled. Here in 20 years, we have 100 X.

88
00:11:52,580 --> 00:11:56,450
But that's only half the story. And it's quite a powerful one already.

89
00:11:56,780 --> 00:12:03,790
The other half, the story is denoted by the different colours. The light pink denotes analogue data.

90
00:12:03,800 --> 00:12:06,680
The dark purple denotes digital data.

91
00:12:06,920 --> 00:12:18,950
And if you look at this white vertical line here, that is the year 2000, the year 2000, three quarters of the data in the world was still analogue.

92
00:12:21,610 --> 00:12:24,620
Today. It's less than 1%.

93
00:12:25,700 --> 00:12:30,860
Within 15 years, we have moved from an analogue world to a digital world.

94
00:12:31,250 --> 00:12:38,210
And that, of course, means that it is easier to collect, easier to store, easier to analyse, easier to retrieve.

95
00:12:39,240 --> 00:12:42,940
What does this do? How can we imagine?

96
00:12:43,090 --> 00:12:47,200
What are the results of this? The consequences? Well, think about it.

97
00:12:47,560 --> 00:12:58,810
The real element, the essence here that I want to convey is that if you increase something radically in quantity, it can take on a new quality.

98
00:13:00,190 --> 00:13:04,490
Consider photography. If I take a photo of a rider on a horse.

99
00:13:04,640 --> 00:13:09,590
That's a photo on the rider. On a horse. If I take a photo every second of a rider on a horse.

100
00:13:09,770 --> 00:13:21,230
That's a lot of photos. All right. On a horse. But what if I take 16 photos per second on a rider on a horse and show them in fast succession?

101
00:13:21,410 --> 00:13:32,630
Then this added quantity of data, a quantity of images that I have, translates into a new quality into in two.

102
00:13:36,220 --> 00:13:45,790
Moving pictures. In essence, what big data provides us with is a new perspective on reality.

103
00:13:47,980 --> 00:13:51,610
Now. How can we characterise that perspective on reality?

104
00:13:51,790 --> 00:13:55,389
Let me try with three words here more messy and correlations.

105
00:13:55,390 --> 00:13:55,960
First, more.

106
00:13:56,200 --> 00:14:05,470
More means that we have more data today available relative to the problem we are studying or the phenomenon that we are trying to investigate.

107
00:14:05,800 --> 00:14:12,130
It's not necessary to have a billion data points. You can have 60,000 data points and still be doing big data.

108
00:14:12,150 --> 00:14:20,170
If if that encapsulates almost all or quite close to all of the phenomenon that you are trying to study.

109
00:14:20,170 --> 00:14:26,770
For instance, if you have that amount of data, that comprehensive dataset available,

110
00:14:26,770 --> 00:14:34,600
then you can let the data speak rather than have it answer questions that you already had in mind when you.

111
00:14:36,130 --> 00:14:42,160
Collected the data. What do I mean with that? Let me use photography again.

112
00:14:42,400 --> 00:14:45,700
This is the moment here that I like the most.

113
00:14:46,150 --> 00:14:50,680
It is the moment when I take a picture of you. Now.

114
00:14:50,680 --> 00:14:54,590
Would you please smile? Okay.

115
00:14:54,860 --> 00:14:57,890
Now, as I take this picture, I have to make a decision.

116
00:14:58,850 --> 00:15:04,520
The decision is, who do I focus on? Do I focus on the Dapper Bill Nighy and the first row?

117
00:15:06,440 --> 00:15:11,330
Or do I focus on you back there and the last row?

118
00:15:12,630 --> 00:15:17,520
If I focus on you, Bill, unfortunately, you get back there, you'll be out of focus.

119
00:15:17,940 --> 00:15:22,470
You'll be blurred. And afterwards, I can bring you back into focus.

120
00:15:23,190 --> 00:15:32,710
It's gone. The data isn't there. So I need to know at the moment of collecting what really is important for me and what isn't.

121
00:15:33,970 --> 00:15:38,400
But what if I don't? Did. Start over again.

122
00:15:38,880 --> 00:15:45,180
Good luck. What if we could have something that would be better than that?

123
00:15:45,780 --> 00:15:48,930
Well, take a photo again as an example.

124
00:15:49,170 --> 00:15:59,630
This is a photo of a toothbrush. It's in focus back in the blurry part of the photo you find, actually.

125
00:16:01,800 --> 00:16:06,930
A photo of my four year old son. I can't put him back into focus.

126
00:16:06,930 --> 00:16:12,350
Right. He's blurred. Too bad to lend a.

127
00:16:13,700 --> 00:16:19,670
This is not a normal photo. This is a photo taken with a big data camera called a lateral light field camera.

128
00:16:19,700 --> 00:16:23,420
Here it is. And so when I take a photo, it's a huge file.

129
00:16:23,570 --> 00:16:27,400
It takes all of the focal pains. It takes all of the light rays in.

130
00:16:27,410 --> 00:16:35,450
And I can click on my son and there he goes and comes into focus because all of the data is present in the photo.

131
00:16:35,720 --> 00:16:37,460
Or I can click, of course, on the toothbrush.

132
00:16:37,760 --> 00:16:46,940
If I'm more interested in that and that comes into focus, I therefore can let the data speak and ask it question.

133
00:16:48,860 --> 00:16:54,860
That I didn't know that I wanted to ask when I collected. The second element is messy.

134
00:16:55,700 --> 00:17:03,100
And let me just very briefly say that in the big data age. We will combine datasets of varying quality.

135
00:17:04,490 --> 00:17:14,300
And that requires us to give up a little bit on our desire coming from the Small Data Age to focus on exactitude of data and data quality.

136
00:17:15,140 --> 00:17:21,740
What we gain in volume, we can trade off in quality, at least to a certain extent.

137
00:17:21,980 --> 00:17:26,700
More and messy together lead to insights through correlations.

138
00:17:26,720 --> 00:17:35,930
Now, in the first class, in statistics, what they tell you is data does not give you causality, only correlations.

139
00:17:37,720 --> 00:17:41,560
That's true. It's incredibly hard for human beings to understand.

140
00:17:42,910 --> 00:17:48,550
So take the example of global supermarket chain Walmart.

141
00:17:49,150 --> 00:17:54,430
Walmart captures all kinds of transactional data about what people buy and when and so forth.

142
00:17:54,430 --> 00:17:56,800
So they get a big data analysis to find out more.

143
00:17:57,220 --> 00:18:07,420
And they discovered that just before a hurricane hits a Wal-Mart location, people go to the Wal-Mart and they buy batteries and flashlights.

144
00:18:09,200 --> 00:18:16,370
Of course I would have thought so. But then they discovered that they also buy Pop Tarts.

145
00:18:19,700 --> 00:18:26,230
Actually strawberry Pop-Tarts. Pop-Tarts is a sugary American snack.

146
00:18:26,240 --> 00:18:32,670
Please note that I do not call it food that is being sold at Walmart.

147
00:18:33,650 --> 00:18:40,469
When they found this out through correlational analysis, these researchers immediately said, Oh my God, why is this the case?

148
00:18:40,470 --> 00:18:44,840
So why is this the case? Oh, and they came up with all kinds of hypotheses, right?

149
00:18:45,080 --> 00:18:48,080
You buy these pop tarts to feed your kids, you buy the Pop-Tarts,

150
00:18:48,080 --> 00:18:54,530
because if you eat a lot of Pop-Tarts, you basically hallucinate the hurricane away or whatever.

151
00:18:56,160 --> 00:19:00,060
Until a researcher said, Stop, timeout.

152
00:19:00,880 --> 00:19:04,750
We don't know. The data doesn't tell us. And guess what?

153
00:19:04,990 --> 00:19:11,200
We don't need to know. All that we need to know is what is happening, not why.

154
00:19:12,070 --> 00:19:16,360
And that is good enough in this particular instance. And so the other said.

155
00:19:16,360 --> 00:19:25,690
That's right. So since then, Wal-Mart before hurricane moved the Pop Tarts from the back of the shop to the front and sells even more of them.

156
00:19:28,150 --> 00:19:29,709
But for us human beings,

157
00:19:29,710 --> 00:19:39,330
that's really hard to understand because we human beings are almost hard coded to see the world as a sequence of of cause and effect.

158
00:19:40,660 --> 00:19:45,309
We can't escape that. Daniel Kahneman, the Nobel laureate in economics,

159
00:19:45,310 --> 00:19:52,630
said that this is this fast thinking that we think we make sense of the world, even though oftentimes.

160
00:19:53,830 --> 00:19:57,100
It's kind of like it makes us feel comfortable.

161
00:19:57,250 --> 00:20:01,690
It gives us the feeling that we understand the world, even though we don't.

162
00:20:02,560 --> 00:20:11,560
And so, for example, when I would have had dinner with you in Hall and one of the colleges that shall remain unnamed here,

163
00:20:11,600 --> 00:20:16,960
just a hypothetical here last night. And I would have had a stomach bug this morning.

164
00:20:18,180 --> 00:20:23,250
I would immediately have connected what I eat in hall with the stomach bug,

165
00:20:23,400 --> 00:20:29,310
even though it is far more likely that I would have gotten my gastroenteritis bug by shaking hands with some of you.

166
00:20:31,510 --> 00:20:35,950
Our brain cannot stop creating these causal linkages.

167
00:20:36,580 --> 00:20:44,140
But when we do so, we must be incredibly careful because often times we are just following the wrong path.

168
00:20:44,170 --> 00:20:53,139
So rather than jumping to a quick conclusion about why things are the way they are, it would be better to just first know what the things are.

169
00:20:53,140 --> 00:20:59,820
They are. Or in other words, we need to learn to walk because we can run.

170
00:21:02,300 --> 00:21:09,910
Hmm. Now. One way of looking at this is through the example of machine translation.

171
00:21:11,350 --> 00:21:21,370
In the 1950s, the US government has had amassed a lot of documents in Russia, but they didn't have enough translators, so they said.

172
00:21:22,640 --> 00:21:28,160
Will translated with the help of computers in the wind, the computer scientists and the computer scientist said, Oh, this is easy.

173
00:21:28,280 --> 00:21:36,020
We teach the 200 or so grammatical rules into the computer, a dictionary to it, and in three months we are done and we have machine translation.

174
00:21:39,310 --> 00:21:44,470
12 years later and about $1,000,000,000 later, that project was declared a failure.

175
00:21:45,430 --> 00:21:51,840
Nothing happened in machine translation until the 1980s, when engineers at IBM had a very different idea.

176
00:21:51,850 --> 00:21:54,970
They said, Why don't we give up on teaching the computer?

177
00:21:55,000 --> 00:22:02,080
Why one word in one language translate into one word in another language and just go with what that is with

178
00:22:02,380 --> 00:22:09,430
statistical correlations and what word is most likely going to be translated from one language into another language?

179
00:22:11,360 --> 00:22:18,950
And they had a great training text that with the proceedings of the Canadian Parliament available in English and French,

180
00:22:18,950 --> 00:22:24,230
and they did it and it worked beautifully. It was the first machine translation to actually was useful.

181
00:22:25,690 --> 00:22:30,429
Then they thought that they could make it even better by. By triggering, by changing.

182
00:22:30,430 --> 00:22:36,630
By tweaking the algorithm. But it didn't really matter very much, so they gave up.

183
00:22:37,200 --> 00:22:40,990
Ten years later, this Start-Up Company in Mountain View that I already mentioned.

184
00:22:41,010 --> 00:22:49,409
Google got into the fray. A German Google engineer by the name of France UX said The problem is not the algorithm.

185
00:22:49,410 --> 00:22:52,800
The problem is the amount of data. We need just more data.

186
00:22:52,980 --> 00:23:00,990
And so they sucked the entire World Wide Web, you know, all the different language versions of the websites of the European Union.

187
00:23:00,990 --> 00:23:10,620
Finally, they're good for something. We're sucked in all of the multilingual websites of these multinational corporations where

188
00:23:11,160 --> 00:23:15,900
all of the PDF users manuals that can be downloaded from your route to do your ironing board.

189
00:23:16,890 --> 00:23:25,920
To your VCR. And when I read this and when I heard this and when friends told it to me, Ken and my and myself, we said, You must be kidding.

190
00:23:26,190 --> 00:23:35,070
I can't even read the English version of the manual of my VCR because it was written in Shenzen or somewhere.

191
00:23:36,800 --> 00:23:44,210
And he said, it doesn't matter if you have so much of it. That little blip in the quality really doesn't matter.

192
00:23:45,390 --> 00:23:53,340
And so machine translation in Google Translate is a wonderful example of more messy and correlations.

193
00:23:54,230 --> 00:23:59,720
But if you think now that this is all just about the Internet and Internet companies, think again.

194
00:24:01,270 --> 00:24:09,040
We heard about health just a little earlier. And health is an area where big data will make tremendous inroads.

195
00:24:10,490 --> 00:24:17,840
Think, for example, of a particularly vulnerable group of human beings, premature babies like this one.

196
00:24:18,680 --> 00:24:25,190
Dr. Callum McGregor at the University Hospital in Toronto had the idea to use big data to help the premature babies.

197
00:24:25,490 --> 00:24:32,120
Premature babies are particularly vulnerable because we discover that they have an infection, often too late.

198
00:24:32,180 --> 00:24:35,670
Symptoms manifest themselves too late. So what did they do?

199
00:24:35,690 --> 00:24:45,140
They had digital sensors that measured the vital signs of these premature babies to the tune of 1200 data points a second.

200
00:24:46,120 --> 00:24:50,919
And they collected them over hours and days and weeks and dozens of babies and then look for

201
00:24:50,920 --> 00:24:55,960
patterns with a high degree of likelihood would predict the onset of an infection later on.

202
00:24:56,350 --> 00:24:57,340
And they found the pattern.

203
00:24:57,820 --> 00:25:05,860
And so they now can predict the likely onset of an infection 24 hours before the symptoms manifest themselves, just by looking at the patterns.

204
00:25:06,130 --> 00:25:10,390
And guess what? There are two kickers in the story. First, what is the pattern?

205
00:25:11,170 --> 00:25:14,290
The pattern is that suddenly the vital signs stabilise.

206
00:25:15,140 --> 00:25:18,230
Rather than going haywire. Who would have thought of that?

207
00:25:18,830 --> 00:25:24,670
You know, every doctor that I ask would tells me if the vital signs stabilise, it's time to go home.

208
00:25:24,680 --> 00:25:28,760
Patients are doing well. It's exactly the other way around with premature babies.

209
00:25:29,270 --> 00:25:33,829
And the other kicker is Dr. Callum McGregor, who is saving babies.

210
00:25:33,830 --> 00:25:38,510
They're. Her doctorate is not in medicine.

211
00:25:38,720 --> 00:25:45,820
She's a computer scientist. So what this tells us is that we need to be humble.

212
00:25:47,850 --> 00:25:53,510
Vis a vis the world around us. Because we understand less than we think we understand.

213
00:25:54,980 --> 00:26:01,460
And we need to take this on board and we need to see whether we can let the data speak in order to understand the world.

214
00:26:02,340 --> 00:26:09,070
With more of its complexity. Now go back to the Google flu trends prediction.

215
00:26:09,700 --> 00:26:17,470
You may have heard in the in the media that Google flu trends were off, were wrong in December of 2012.

216
00:26:17,710 --> 00:26:23,470
Here on the very right hand side, you see the spike. Google flu trends predicted more flu cases than there actually were.

217
00:26:24,540 --> 00:26:35,160
What was going on. What was going on was actually something that is crucial and brings in two Europeans, Mr. Bull and Mr. Bayes.

218
00:26:37,040 --> 00:26:43,130
Because if you ask somebody, what's the chance when you throw a coin that it lands, heads up.

219
00:26:44,280 --> 00:26:51,120
Most people will tell you 50%. That's a really good approximation, but it is actually wrong.

220
00:26:52,720 --> 00:26:57,670
Because if you throw a particular coin, it's not 5050.

221
00:26:57,970 --> 00:27:02,110
It never is. Everybody throws slightly differently. Every coin is slightly different.

222
00:27:05,230 --> 00:27:08,320
We just have an approximation 5050 works pretty okay.

223
00:27:08,590 --> 00:27:15,760
Like Newton's gravitation law was pretty okay until we needed to have GPS and then we need it to go beyond that.

224
00:27:15,760 --> 00:27:24,219
But. But it's the same here. And so we need to whenever we have more data available and new data available,

225
00:27:24,220 --> 00:27:28,900
feed that in and rethink our view of the world, our perspective of the world.

226
00:27:29,530 --> 00:27:36,610
That's why the Bayesian idea of using priors is so fundamental to what big data is all about.

227
00:27:36,760 --> 00:27:43,239
And the mistake that Google made is that they once developed a model for flu trends in 2009 and

228
00:27:43,240 --> 00:27:49,389
stuck to it and never updating it based on the additional data that it had when they updated it.

229
00:27:49,390 --> 00:27:54,610
With 2010 and 2011 data, their December 2012 forecast was almost spot on.

230
00:27:57,580 --> 00:28:03,400
So we need to abandon our idealised worlds and approximating reality through big data.

231
00:28:04,090 --> 00:28:08,650
No. This means.

232
00:28:09,610 --> 00:28:20,290
That what is happening here at the Mathematical Institute is core is essential because we need to develop almost a new language,

233
00:28:20,290 --> 00:28:24,160
a new way of thinking, a new toolset for this big data world.

234
00:28:25,510 --> 00:28:27,850
And you are doing it here.

235
00:28:29,610 --> 00:28:37,470
And when we look at it, most of the tools that we have in the small data world of making sense of data, you know, like our square.

236
00:28:38,250 --> 00:28:46,470
Most of the software tools that we have like are developed here at Oxford were originally developed in a small data context.

237
00:28:47,700 --> 00:29:00,060
Now we need to find their equivalents in the big data. And that is, as I understand it, with the Oxford Nye NIE Lab will provide us.

238
00:29:01,530 --> 00:29:11,040
Cutting edge research and trying to help us enter this big data world with the right tools so that we can make sense of it.

239
00:29:11,910 --> 00:29:17,880
For that. Of course, we need data ification, that is to render every more aspect of our world into data form.

240
00:29:18,060 --> 00:29:19,890
We all need to do that location, right?

241
00:29:20,640 --> 00:29:28,020
I still remember that when you would take a car and drive into a new city and somebody would have a map on the lap sitting next to you.

242
00:29:29,600 --> 00:29:32,600
None of my students here in the audience ever would do that anymore.

243
00:29:33,680 --> 00:29:37,430
You did that? They asked me. Didn't you have sat, NAV?

244
00:29:38,180 --> 00:29:44,200
No. Location has been data fied and therefore we can analyse it.

245
00:29:44,800 --> 00:29:53,300
But it's not just that. In Japan, researchers have data five human behinds through a 30 sensors.

246
00:29:53,420 --> 00:29:59,330
They measured the size of the bum. Why?

247
00:30:01,920 --> 00:30:06,350
Never ask researchers that question. Why do they do that?

248
00:30:06,740 --> 00:30:11,710
Because they think. That every person has a different bum.

249
00:30:11,950 --> 00:30:15,730
They discover that the bum is as unique as a fingerprint.

250
00:30:16,180 --> 00:30:20,290
And so the idea is to use this as a car anti-theft device.

251
00:30:22,850 --> 00:30:26,510
You get into your car. Your car measures your bum.

252
00:30:27,080 --> 00:30:32,660
You can drive off. The thief gets into the car.

253
00:30:33,440 --> 00:30:39,160
Bum is measured. Bum is way too big. Car stops and the tracks.

254
00:30:40,430 --> 00:30:43,550
That's the data ification of another aspect of our reality.

255
00:30:43,670 --> 00:30:46,130
Now, you know, of course, this this is Google Glasses.

256
00:30:46,640 --> 00:30:52,280
In this version, Google Glass doesn't do what I think Google Glass is going to do and where Google is investing in it,

257
00:30:52,280 --> 00:30:56,180
and that is to data for the human gaze. What are we looking at?

258
00:30:56,510 --> 00:31:03,170
Can you imagine how valuable it is to know what people, what human beings are looking at like, what advertisements they are looking at,

259
00:31:03,410 --> 00:31:07,850
what are they looking at in the shopping window, what men are looking at when they walk down the street?

260
00:31:08,600 --> 00:31:12,960
Well, skip the last one. We know that. Cars.

261
00:31:15,920 --> 00:31:23,930
Data ification that is rendering ever more aspects of our reality into data form permits us to then extract value.

262
00:31:23,930 --> 00:31:33,170
And if you take anything about the value proposition of big data away from this talk, please take away the fact that in the small data rich, we.

263
00:31:34,630 --> 00:31:39,010
Use data for a particular purpose for which it was collected, and then we threw it away.

264
00:31:39,850 --> 00:31:45,850
And the big data age that we understand that the value of data is not exhaustive by using it once,

265
00:31:46,180 --> 00:31:50,050
but we can reuse it and reuse it and reuse it and reuse it for multiple purposes.

266
00:31:50,890 --> 00:31:54,370
Extracting more and more data, much like a iceberg.

267
00:31:55,350 --> 00:32:02,670
With data. Most of the latent value has been untapped so far.

268
00:32:02,850 --> 00:32:12,930
But we can tap into it to reuse. For example, global financial payment company Swift, that is transferring money across borders.

269
00:32:13,470 --> 00:32:21,990
Swift discovered that it can use its data to predict the health of local economies because of the correlation that it found.

270
00:32:22,560 --> 00:32:26,639
That's the re-use of data that it has. Or Start-Up Company.

271
00:32:26,640 --> 00:32:37,680
INRIX in Seattle helps over 100 million users every single working day to find their way to work or back with their car around traffic jams.

272
00:32:37,860 --> 00:32:41,310
Creating heat maps like this. Telling them where there is heavy traffic.

273
00:32:41,910 --> 00:32:46,270
Now. SAT maps do that right.

274
00:32:46,960 --> 00:32:52,000
But my satnav is stupid. My satnav tells me that there is a heavy traffic.

275
00:32:53,030 --> 00:32:58,850
When I'm already in it. This knows when heavy traffic is forming.

276
00:32:59,090 --> 00:33:02,750
Why? How? Because they have the data. How do they get the data?

277
00:33:03,320 --> 00:33:11,479
Because every one of their 100 million users is a sensor sending back data on where they are, how fast they are going.

278
00:33:11,480 --> 00:33:15,380
And so. And then, you know what? INRIX found out?

279
00:33:15,560 --> 00:33:22,459
That they can reuse the data. And they teamed up with a hedge fund because it turns out that there is a correlation between heavy

280
00:33:22,460 --> 00:33:29,420
traffic on weekends around shopping malls and revenue numbers of the shops in the shopping malls.

281
00:33:29,720 --> 00:33:35,390
So they are buying or selling stocks before quarterly earnings results based on the predictions.

282
00:33:35,630 --> 00:33:42,290
That, too, is a reuse of data. And when you have so much reuse happening, you can rethink your business model.

283
00:33:42,800 --> 00:33:47,540
Take Royce Weiss, not the luxury car company, but the jet engine producer.

284
00:33:47,540 --> 00:33:51,569
Number two, world's jet engine producer. Royce.

285
00:33:51,570 --> 00:33:55,590
Royce used to produce and sell jet engines.

286
00:33:55,740 --> 00:33:59,670
These jet engines have lots of sensors in there that measure vibration and temperature and

287
00:33:59,670 --> 00:34:04,680
pressure and so forth and send it to a computer in the jet engine that manages the jet engine.

288
00:34:05,310 --> 00:34:08,850
And then the data are thrown away with the Airbus 380 engine here.

289
00:34:08,850 --> 00:34:17,820
You see, when they discover that they can actually capture the data and then send it back to Royce Royce headquarters once the plane has landed.

290
00:34:18,720 --> 00:34:24,810
They do that. It's an enormous amount of data, a couple of gigabytes per plane, per plane ride.

291
00:34:25,710 --> 00:34:33,930
But what do they do with it? They do an analysis to find patterns that show them when a part in the jet engine is breaking before it actually breaks.

292
00:34:34,950 --> 00:34:43,229
So they can do what is called predictive maintenance, which is great because you can do them the maintenance before the part breaks.

293
00:34:43,230 --> 00:34:46,350
That is when the plane is on the ground. That helps,

294
00:34:46,860 --> 00:34:53,969
but it also helped tremendously Royce West to change its business model around from a company selling

295
00:34:53,970 --> 00:35:00,510
jet engines to a company selling fixed fee maintenance contracts and going into the service sector.

296
00:35:00,720 --> 00:35:05,250
Today, they are now 70% of their revenue derived from services.

297
00:35:06,200 --> 00:35:12,710
Many of you will now look at this and say this means only that the big get bigger, the Googles, the Facebooks and so forth.

298
00:35:13,040 --> 00:35:19,309
And there is some truth to it because the big ones are buying up data ingestion platforms.

299
00:35:19,310 --> 00:35:25,790
Remember that Google bought a thermostat company earlier this year called Nest for almost $3 billion.

300
00:35:26,120 --> 00:35:29,360
A thermostat company. Give me a break. Thermostat company?

301
00:35:30,730 --> 00:35:32,930
But Google didn't buy a thermostat company.

302
00:35:32,950 --> 00:35:43,030
Google bought a data ingestion platform that collects data about how cold or warm people want their rooms in their houses to be.

303
00:35:43,420 --> 00:35:46,600
And that was worth, in their mind, $3 billion.

304
00:35:46,960 --> 00:35:58,000
So. So big data we will have in this context in the financial services industry to improve forecast and decision making in the industry.

305
00:35:58,030 --> 00:35:59,320
We already see that happening.

306
00:35:59,650 --> 00:36:07,750
We also have it to stimulate innovation generally because with big data there will be many businesses coming up with many new ideas.

307
00:36:08,710 --> 00:36:22,150
But also it will make the financial information sector itself into a data platform, which is precisely what Bill NI and FTT have been doing.

308
00:36:23,020 --> 00:36:33,040
So it's not just that the big get bigger, but that there is a place for Start-Up companies, for vibrant new entrepreneurs.

309
00:36:33,790 --> 00:36:41,440
Think of the site dot com a company that predicts whether for 50,000 consumer goods, what are the prices going up or down?

310
00:36:41,530 --> 00:36:45,520
And they are so convinced that their prediction is right that if you buy the product based on their prediction.

311
00:36:47,210 --> 00:36:49,400
And it's wrong. They will refund you the difference.

312
00:36:51,600 --> 00:37:01,080
This is the brainchild of Oren Etzioni, and they do have the desired outcome, had hundreds of thousands of customers every day,

313
00:37:01,140 --> 00:37:06,360
billions of data points in order to do the calculation of the prediction for 50,000 consumer products.

314
00:37:06,810 --> 00:37:12,430
You know how many employees they have. 30, including the cleaning lady.

315
00:37:14,050 --> 00:37:21,330
How many servers? Zero because they all do it in the cloud.

316
00:37:22,600 --> 00:37:27,580
And that means that there isn't a huge investment anymore necessary to start this up.

317
00:37:27,760 --> 00:37:36,670
So we'll see a lot of activity on the small end of the spectrum as well, with nimble companies being extremely successful.

318
00:37:37,920 --> 00:37:43,080
Now, let me briefly tell you that this also changes how we internally operate.

319
00:37:45,670 --> 00:37:50,290
In organisations. I'm not talking about 50 Shades of Grey.

320
00:37:50,710 --> 00:37:59,740
I'm talking about 41 Shades of Blue. That is a few years back, Google had to decide what kind of colour to use for its Google search box.

321
00:38:01,510 --> 00:38:06,820
And the designer used a particular colour and his supervisor said, Why does colour?

322
00:38:06,860 --> 00:38:11,530
And he said, Because I'm the designer. And she said, Did you do a test?

323
00:38:11,530 --> 00:38:15,610
And he said, No. If you want to do me a test, I will resign. She said, Resignation accepted.

324
00:38:17,410 --> 00:38:23,530
And then they did a test. And then they found out testing 41 different shades of blue, that there was a slightly different shades of blue.

325
00:38:24,530 --> 00:38:31,100
That worked better. And gave Google $12 million more in annual revenues ad revenues per year.

326
00:38:32,580 --> 00:38:38,639
So the person who fired that, chief designer, Marissa mayer, now chief of Yahoo!

327
00:38:38,640 --> 00:38:49,050
Said this was the best business decision she ever made. It means that self-styled experts will be questioning.

328
00:38:50,850 --> 00:38:54,450
What's your email? What's the factual basis?

329
00:38:54,750 --> 00:38:59,670
They will be asked for your assumptions, for your suggestions.

330
00:39:01,270 --> 00:39:06,730
You know, and that especially is valid in the financial services industry.

331
00:39:07,180 --> 00:39:17,020
Companies like Imaginative in the Silicon Valley are looking at creating interesting trading platforms for specialised products.

332
00:39:17,020 --> 00:39:24,370
But then there is FTT, which creates with its big data set,

333
00:39:24,370 --> 00:39:30,880
not just a platform for people to learn, but for those to teach to people what they are learning.

334
00:39:31,300 --> 00:39:40,760
Learn as well, learn about the learning, and therefore create an environment that keeps people engaged and interested.

335
00:39:40,850 --> 00:39:45,960
Just amazing. But then I'm sure you have also heard of Target.

336
00:39:47,150 --> 00:39:54,530
This superstore in the United States that was able to predict with a relatively good degree of likelihood.

337
00:39:56,080 --> 00:40:02,320
Based on transactions, shopping transactions that one of their customers was pregnant.

338
00:40:03,650 --> 00:40:09,290
Even though that customer might not herself know if the pregnancy and.

339
00:40:11,590 --> 00:40:20,970
Oh. And so there is a dark side of big data, too, and we need to mention that.

340
00:40:22,160 --> 00:40:26,330
Many of you now think of George Orwell's 1984 and Surveillance Society.

341
00:40:26,330 --> 00:40:31,680
Yes, yes, yes, yes. Snowden, I get it. But this is only half the problem.

342
00:40:31,740 --> 00:40:40,410
The other half is that with predictions, we are also able perhaps to predict human behaviour.

343
00:40:41,380 --> 00:40:49,630
And as we predict human behaviour, we might be tempted to hold people responsible not for what they have done but what they are only predicted to do.

344
00:40:50,440 --> 00:40:56,170
Now if you think of the Hollywood movie Minority Report, that's precisely what I'm thinking of, too.

345
00:40:56,650 --> 00:41:04,570
In 30 states, in the United States, the decision of whether or not you are coming free out of prison on a parole.

346
00:41:05,610 --> 00:41:12,510
Is being made in part by a big data algorithm that is predicting whether or not you're going to be a criminal in the next 12 months.

347
00:41:13,800 --> 00:41:20,650
Is this going to be the end of free will? Because everything will be predicted about human behaviour.

348
00:41:23,950 --> 00:41:30,940
We need to be very aware of those ethical dilemmas, but we need to also keep in mind that the problem here is not big data.

349
00:41:32,050 --> 00:41:36,040
The problem is how we use the results of big data.

350
00:41:36,490 --> 00:41:42,459
Correlations are not causality using correlational insights for causal purposes.

351
00:41:42,460 --> 00:41:46,570
For example, to assign responsibility is abusing big data.

352
00:41:47,260 --> 00:41:56,110
It's stupid, it's wrong. You know, in the United States, they did a big data analysis to decide what kind of car has the least repairs,

353
00:41:56,110 --> 00:42:09,860
and it turns out it's a car of the colour orange. Half of you are already thinking why?

354
00:42:10,370 --> 00:42:13,190
Right. Is it because an orange car is more visible at night?

355
00:42:13,340 --> 00:42:17,900
Is it because our owner of an orange cause drives more carefully because it's a special car?

356
00:42:18,170 --> 00:42:24,260
Is it because it was manufactured specially? Timeout guys and gals, time out.

357
00:42:24,470 --> 00:42:36,110
The data doesn't tell us. The moment you start imbuing data with more meaning than it has that moment you succumb to the dictatorship of data.

358
00:42:38,210 --> 00:42:42,260
And we need to be aware of that. So let me come to a conclusion.

359
00:42:43,820 --> 00:42:47,420
Big data is going to change how we make decisions, how we live,

360
00:42:47,420 --> 00:42:55,400
work and think from how we learn to what kind of medications we are getting to how cars drive themselves.

361
00:42:56,210 --> 00:42:59,340
But big data also has. A number of challenges.

362
00:42:59,370 --> 00:43:06,659
So what is really important is as we enter this big data age that we where we remain fully in control of that technology,

363
00:43:06,660 --> 00:43:09,810
aware of its constraints and its limitations.

364
00:43:11,320 --> 00:43:19,710
That. As much as. We will utilise the insights from big data.

365
00:43:19,890 --> 00:43:24,990
We also protect and preserve a space for the human.

366
00:43:26,420 --> 00:43:35,309
For creativity, originality. For irrationality, for sometimes acting in defiance of what the data says.

367
00:43:35,310 --> 00:43:41,320
Because at the end of the day. Data is just a shadow of reality.

368
00:43:42,490 --> 00:43:47,630
And therefore it is always incomplete. And always a little bit incorrect.

369
00:43:48,080 --> 00:43:55,130
And so we need to approach this new big data world with a lot of humility rather than hubris.

370
00:43:56,000 --> 00:43:59,270
And we need to do so with a lot of humanity.

371
00:43:59,840 --> 00:44:00,710
Thanks very much.