1
00:00:00,030 --> 00:00:06,599
Why don't we begin? It is a pleasure today to introduce our speaker, Professor Chris Lynn.

2
00:00:06,600 --> 00:00:10,710
Talk to our own sub Department of Astrophysics.

3
00:00:10,830 --> 00:00:18,870
Chris was a an undergraduate at Cambridge University and did his graduate studies at UCL.

4
00:00:19,830 --> 00:00:29,940
The topic of his PhD thesis was essentially the chemistry of sulphur in the context of star formation.

5
00:00:30,600 --> 00:00:45,890
So when I heard this, I couldn't resist putting to him that between the study of brimstone and the observations of the fires of the first star,

6
00:00:45,900 --> 00:00:50,280
was he not in fact the very incarnation of Satan's astronomer self?

7
00:00:51,870 --> 00:01:01,829
Professor Lintott did not deny the fact that he had moved on from his interest in study for us excuse me,

8
00:01:01,830 --> 00:01:08,370
in his interest in star formation turning into an interest in cosmological star formation.

9
00:01:08,850 --> 00:01:14,070
Led him to his current field of interest, which is galaxies.

10
00:01:14,820 --> 00:01:26,550
And Chris became interested in the problem, which we'll hear about today, which is what is the best way to find peculiar morphology in galaxies?

11
00:01:26,580 --> 00:01:30,420
This is something that a computer by itself is not particularly good at.

12
00:01:31,020 --> 00:01:36,150
For all of its speed, because peculiar is hard to quantify.

13
00:01:37,620 --> 00:01:43,110
Human beings can't look through millions of galaxies by themselves unaided.

14
00:01:43,530 --> 00:01:54,780
So the question is, how do you put these two resources together to try and optimise the search using human brains with the aid of computer technology?

15
00:01:55,230 --> 00:02:05,400
And in that respect, he has succeeded brilliantly, establishing a prototype program for not just galaxies,

16
00:02:05,400 --> 00:02:09,660
but citizen science in general and its use in several different fields.

17
00:02:10,170 --> 00:02:20,330
So he is going to tell us today about citizen science, the story of Zooniverse from Galaxy Zoo to ALS City.

18
00:02:21,150 --> 00:02:24,450
And we get to begin with the penguin. Excellent.

19
00:02:24,930 --> 00:02:28,530
Thank you. Cheers. Well, thank you for that introduction.

20
00:02:29,970 --> 00:02:33,210
Satan's astronomer. Sounds beautifully well-paced, doesn't it?

21
00:02:33,630 --> 00:02:37,410
Not sure. What? Not sure what overheads you get on one soul.

22
00:02:37,440 --> 00:02:41,010
But I'm sure finance will let us know before the end of the lecture.

23
00:02:42,060 --> 00:02:45,150
Thank you. It's a delight to be asked to give the colloquium.

24
00:02:46,200 --> 00:02:51,410
I am assuming that I was selected because I speak loudly enough that I can keep you awake on a Friday afternoon.

25
00:02:51,420 --> 00:02:56,530
So I shall try to do that. And there will indeed be more pictures of penguins later in the talk.

26
00:02:56,550 --> 00:03:02,640
This is actually a peculiar galaxy not found by but adopted by some of the volunteers.

27
00:03:03,420 --> 00:03:08,310
But I thought I'd begin by pointing out how modern astronomy is done.

28
00:03:08,580 --> 00:03:16,680
And I was wrong about this for almost all of my life. I'm one of the few, I think, professional astronomers who grew up as amateur astronomers.

29
00:03:16,680 --> 00:03:22,620
I had a small telescope and from an early age I was thrilled with the idea that I might make a discovery.

30
00:03:23,460 --> 00:03:30,750
Ideally, I wanted to discover a comet because that named after discovers Comet Lintott I think has a particularly fine ring,

31
00:03:31,800 --> 00:03:40,260
but I would have taken anything and equipped with a six inch reflector and observing from the light polluted skies of suburban South Devon,

32
00:03:41,100 --> 00:03:46,440
I found I stood a pretty good chance. The closest I ever got was in viewing this thing.

33
00:03:46,440 --> 00:03:47,489
The Nebulosity.

34
00:03:47,490 --> 00:03:54,000
The gas you see at the bottom of this image is part of the Orion Nebula, a great star forming complex that's visible in the winter sky.

35
00:03:54,300 --> 00:03:59,220
And one evening I nudged my telescope, wasn't equipped with a drive, so the thing just moved.

36
00:03:59,460 --> 00:04:04,650
And this star cluster came into view. I looked it up in my star atlas.

37
00:04:05,550 --> 00:04:11,879
This was before we had the Internet at home, so I had to look at a book for the younger members of audience.

38
00:04:11,880 --> 00:04:17,580
With the audience, the book is like the internet, except that you have to turn the pages manually.

39
00:04:18,180 --> 00:04:20,930
And this cluster wasn't in the atlas,

40
00:04:20,940 --> 00:04:32,219
and I remember getting a pencil and marking across and putting Lyn top one next to now I the thing I like about this idea is

41
00:04:32,220 --> 00:04:38,160
that I think my scientific tendencies are clear in the fact that I used a pencil because this is a provisional result and this.

42
00:04:38,160 --> 00:04:44,309
Lintott One turns out to be better known as NGC 1981, though of course it's still the top one.

43
00:04:44,310 --> 00:04:50,100
To me, it's a perfectly ordinary cluster that was discovered in the 19th century, but looks rather nice at a small telescope.

44
00:04:50,430 --> 00:04:53,739
And this I think, was the start of my growing up as an astronomer.

45
00:04:53,740 --> 00:04:59,760
And I realised that the days in which discoveries were made by ordinary people with small telescopes had long.

46
00:04:59,980 --> 00:05:04,450
And telescopes these days look more like this.

47
00:05:04,750 --> 00:05:10,930
This is the Sloan Digital Sky Survey Telescope that if you, of course, now in New Mexico, a place that like Oxford,

48
00:05:10,930 --> 00:05:21,190
has about 300 clear nights a year and to first order and at Sloan is I think is what happens

49
00:05:21,190 --> 00:05:25,330
when you let particle physicists build your astronomy experiment because it's an experiment.

50
00:05:25,810 --> 00:05:27,790
It was designed not as an observatory,

51
00:05:28,150 --> 00:05:35,350
not as a place that you visit to take control of the telescope and point out your favourite targets, but as a data production machine.

52
00:05:35,710 --> 00:05:42,280
And Sloan allowed the sky to turn over it for eight years, measuring the position of more than 300 million objects.

53
00:05:42,640 --> 00:05:47,320
Of those, 8 million of them were identified as faint, fuzzy things, probably galaxies.

54
00:05:47,650 --> 00:05:53,760
And on the clearest and still is nights, Sloan would return to those and use spectroscopy to measure a distance.

55
00:05:53,770 --> 00:05:57,610
So this thing is making a three dimensional map of our local universe,

56
00:05:57,610 --> 00:06:03,580
and it's doing that so that we can make rather crude comparisons to sophisticated cosmological models

57
00:06:03,880 --> 00:06:08,140
because we want to understand the physics that's driven the large scale evolution of the universe.

58
00:06:08,980 --> 00:06:16,480
So, for example, we take our eight years worth of data on how many galaxies and we reduce them to a mass function.

59
00:06:16,870 --> 00:06:23,980
So this is the density of galaxies in the in the Sloan volume of particular stellar masses.

60
00:06:24,730 --> 00:06:29,080
So the dots here are data black from Sloan Blue for other surveys.

61
00:06:29,350 --> 00:06:34,510
And the green and the red are different flavours of large scale cosmological simulations.

62
00:06:34,980 --> 00:06:38,500
And for those who aren't astronomers, this is a good fit to the data.

63
00:06:39,460 --> 00:06:45,070
We consider ourselves well satisfied with this. But we, of course, have all sorts of interesting properties.

64
00:06:45,070 --> 00:06:50,800
You can see that there's a problem about high mass where we over predict the number of large galaxies.

65
00:06:51,100 --> 00:06:57,310
And you could see there's a problem at low mass as well, where while you can take your picture, you can over, over or under predict.

66
00:06:58,090 --> 00:07:01,210
And if you're a theorist, this is, of course, a problem with the observations.

67
00:07:01,840 --> 00:07:05,890
And if you're an observer, then you go away, you redo the model.

68
00:07:06,520 --> 00:07:18,150
But that seems something slightly reductive about reducing galaxies to point particles that trace the underlying evolution of the universe.

69
00:07:18,160 --> 00:07:24,130
You can do that. You could have particles with particular mass, of course, and density and and perhaps a size.

70
00:07:24,970 --> 00:07:31,810
But when you look at these things and these are all galaxies drawn from Sloan, admittedly local ones with roughly the same mass,

71
00:07:32,140 --> 00:07:37,719
you see that there's more information here and we can call it morphology if you want to sound sophisticated.

72
00:07:37,720 --> 00:07:43,630
But the shape of the galaxy encodes its integrated dynamical history.

73
00:07:44,050 --> 00:07:49,240
The shape, as James Binney and others will tell you at great length and with great certitude, great understanding.

74
00:07:49,540 --> 00:07:52,390
And the shape of the galaxy depends on the orbits of the stars.

75
00:07:52,690 --> 00:07:57,910
The orbits of the stars are a measure of the dynamical environments through which the galaxies have passed.

76
00:07:58,270 --> 00:08:05,380
And so if you know the shape of the galaxy, you can say something about how it's interacted with its surroundings,

77
00:08:05,620 --> 00:08:10,120
how it's interacted with other galaxies, and even how and where it stars have formed.

78
00:08:10,480 --> 00:08:18,100
And this is not a new idea. Edwin Hubble was the first to sort of systematically look at this in the thirties,

79
00:08:18,370 --> 00:08:25,390
wrote a book called The Realm of the Nebulae, in which he proposed a system for classifying galaxies for Hubble's tuning fork.

80
00:08:25,920 --> 00:08:34,810
And there's some sense that he saw it as an evolutionary scheme with elliptical galaxies like this one collapsing under gravity to form disks,

81
00:08:35,110 --> 00:08:39,430
which then wound up on wound themselves into various types of spiral.

82
00:08:39,790 --> 00:08:46,150
It's not really clear whether even Hubble believed that. But nonetheless, he knew that the shape of these galaxies was important.

83
00:08:46,810 --> 00:08:50,140
And back then, and throughout the fifties and sixties,

84
00:08:50,590 --> 00:08:57,489
we had few enough high resolution images of galaxies that eminent professors would devote themselves to creating, first of all,

85
00:08:57,490 --> 00:09:03,700
new classification schemes, then classifying all the while image galaxies according to those schemes,

86
00:09:03,970 --> 00:09:06,940
and then going to conferences and arguing about whose scheme was best.

87
00:09:07,720 --> 00:09:12,850
And so they multiplied in both a number, the classification schemes, but also in sophistication.

88
00:09:12,860 --> 00:09:21,980
So this is not just the spiral galaxy, this might be it's a spiral galaxy or let's be or it might be an SB three and so on and so forth and so on.

89
00:09:22,480 --> 00:09:29,620
And that kept people happy doing detailed studies of local galaxies for a long time.

90
00:09:29,620 --> 00:09:36,699
But by the 1980s, surveys had improved. And we had things like the Palomar Sky Survey, which I discovered the other day.

91
00:09:36,700 --> 00:09:43,240
To my surprise, this was the first really good survey of the northern sky to go deep,

92
00:09:43,570 --> 00:09:47,920
and it was able to take deep images of the sky because of a new type of photographic plate.

93
00:09:48,520 --> 00:09:52,390
This is in the 1980s. So this is the last pre-digital survey.

94
00:09:53,320 --> 00:09:59,490
Palomar produced thousands of galaxies that were worth classified and simultaneously the astronomical world in the late.

95
00:10:00,160 --> 00:10:04,149
Discovered that you don't need to be a professor to like Galaxy something grad students can do

96
00:10:04,150 --> 00:10:09,010
perfectly good job just as the number of galaxies that needed to be classified reached thousands.

97
00:10:09,370 --> 00:10:14,470
People started that phrase by classifying a thousand galaxies and using that data.

98
00:10:14,710 --> 00:10:20,620
So this worked very well. But people could see that projects like Sloan were coming and that this would it work.

99
00:10:21,190 --> 00:10:24,970
When you have a million galaxies classified, you need to find new approaches.

100
00:10:25,570 --> 00:10:32,470
And there's an excellent paper by my old supervisor with whom I never discussed Galaxy morphology during my Ph.D. offer Laugh.

101
00:10:32,500 --> 00:10:35,680
He's now at UCL when he does two things. First of all,

102
00:10:35,710 --> 00:10:42,280
he gets a panel of experts to classify a set of galaxies and then shows in this in the footnotes that

103
00:10:42,280 --> 00:10:47,860
you can derive who was who student just from their classifications with no other prior knowledge.

104
00:10:48,790 --> 00:10:54,730
But then secondly, he starts the push towards decent machine learning to attack this problem.

105
00:10:54,910 --> 00:11:03,760
And people do. People like Mick Bull, who is at Sussex and then Nottingham do whole PhDs on neural network approaches to galaxy classification.

106
00:11:04,450 --> 00:11:08,440
And almost anyone who attempts this discovers that 70% of the galaxies are easy.

107
00:11:09,250 --> 00:11:14,590
Getting to 80% takes a lot of effort, and beyond that you're really stuck and your accuracy is important.

108
00:11:14,590 --> 00:11:18,010
And we'd like to have reliable cross-sections for the vast majority of these galaxies.

109
00:11:18,010 --> 00:11:23,350
Or at least we'd like a system that will tell us which galaxies are reliably classified.

110
00:11:23,800 --> 00:11:28,210
The machine learning approaches of the time can't even tell you which 30% are wrong.

111
00:11:29,440 --> 00:11:40,240
When I arrived in Oxford, a new solution had been found, which was to to find students who were more committed to the subject.

112
00:11:41,020 --> 00:11:44,340
I think you particularly if a guy for coverage, Lewinsky, who's now 88.

113
00:11:45,010 --> 00:11:55,230
As a senior fellow there and Kevin had looked at 50,000 galaxies and he'd shown, amongst other things, first of all, that that's the limit.

114
00:11:55,240 --> 00:11:59,920
A student will look at 50,000 galaxies before he tells you where to stuff the rest of them.

115
00:12:00,970 --> 00:12:08,830
I'm paraphrasing so we can think of this as the Kevin Limit and other students approach, but do not exceed the Kevin limit.

116
00:12:09,040 --> 00:12:13,420
And then the other thing is that he showed really that it mattered to have a person look at this.

117
00:12:13,870 --> 00:12:17,080
His consultations with different from those you obtain from your networks.

118
00:12:17,350 --> 00:12:22,360
And when you chased the discrepancies down, you found that Kevin was right, at least according to the experts.

119
00:12:25,540 --> 00:12:33,190
I often joke that the obvious solution is to have 20 Ph.D. students work on this, but even then you still only got one classification per galaxy.

120
00:12:33,550 --> 00:12:38,170
And you're you have this uncomfortable truth that any result you produce depends on these classifications.

121
00:12:38,170 --> 00:12:41,590
You have no way of checking that. And so a broader approach was necessary.

122
00:12:41,920 --> 00:12:46,180
And without really thinking about it, we got some friends to knock together a website for Galaxy Z.

123
00:12:46,420 --> 00:12:51,819
This is what it looked like. It gave you an image of the galaxy and it gave you six buttons are not generic.

124
00:12:51,820 --> 00:12:57,730
It's a elliptical galaxy. We if it's a spiral galaxy, if it's a spiral, we want to know which way the arms are going.

125
00:12:58,090 --> 00:13:02,319
You want to hear that story? Ask me about it. In the questions we ask for mergers.

126
00:13:02,320 --> 00:13:06,640
And then a few stars got through and we put the song online.

127
00:13:07,360 --> 00:13:15,130
Pete Wilson and friends in the University Press Office helped us get some attention from the BBC, and within a day we were doing 1.2.

128
00:13:15,160 --> 00:13:20,230
Kevin Weeks and our own Kevin Week is the unit of Galaxy Classifications and 50,000.

129
00:13:20,380 --> 00:13:23,020
We were doing something like 70,000 classifications an hour.

130
00:13:24,340 --> 00:13:30,190
We haven't kept going at that rate, but Galaxy Series received literally hundreds of millions of classifications of galaxies.

131
00:13:30,400 --> 00:13:34,600
And the impressive thing is that taken collectively, those classifications look very good.

132
00:13:35,350 --> 00:13:40,600
So for each of these galaxies, I not only can sort them into their categories, but I can say so.

133
00:13:40,600 --> 00:13:46,720
This one, it's an elliptical. And I have a measure of confidence because ten out of ten people said it was a little sceptical,

134
00:13:46,990 --> 00:13:52,140
whereas this galaxy is a spiral, but only seven out of ten people say it's a spiral.

135
00:13:52,150 --> 00:13:56,530
So we have this not only a classification but an estimate of accuracy.

136
00:13:56,770 --> 00:14:02,530
That number, that vote fraction is proportional to the probability that this really is a spiral.

137
00:14:03,190 --> 00:14:10,480
And so that data for the first time gives us a really powerful, reliable, consistent set of morphologies.

138
00:14:11,110 --> 00:14:15,520
And we went on Galaxy Zoo originally was designed to do the simplest thing that was useful.

139
00:14:15,700 --> 00:14:18,750
Is it elliptical or is it a spiral these days? Yeah.

140
00:14:18,780 --> 00:14:23,650
The Galaxy Zoo. You will find a complicated decision tree that still requires no prior knowledge,

141
00:14:23,920 --> 00:14:29,050
but which takes you through the detailed is there are bulge, there is what shape of a spiral arms?

142
00:14:29,350 --> 00:14:37,270
If so, how many are there? And so on and so forth. Each piece of information encoding a different part of the galaxies dynamical past.

143
00:14:38,110 --> 00:14:44,020
And you also find that we're no longer restricted to Sloan. We quickly finished classifying all of the Sloan galaxies,

144
00:14:44,320 --> 00:14:50,530
and we've now branched out to most of the large surveys that the Hubble Space Telescope has done.

145
00:14:50,530 --> 00:14:59,470
So we're able to compare local morphology, the population of galaxies we see around us with those that existed, perhaps at a redshift of one.

146
00:14:59,590 --> 00:15:01,510
A few billion years ago also.

147
00:15:01,510 --> 00:15:08,200
And this is I don't have time to talk about this in detail, but for example, if you're interested in whether galaxies have a bar at the centre,

148
00:15:08,200 --> 00:15:17,019
this linear feature you can see in this plot, the grey is slightly poorly conceived simulation work based on a small number of simulations.

149
00:15:17,020 --> 00:15:23,229
But look at the data. This is the fraction of spiral galaxies that have a bar as a function of redshift and in

150
00:15:23,230 --> 00:15:27,940
pink from Tom Melvin in Portsmouth and then in black for Brook Simmons here in Oxford,

151
00:15:28,360 --> 00:15:33,459
we can show this this decline in the fraction of barred galaxies over time.

152
00:15:33,460 --> 00:15:38,110
And this is interesting because various people have predicted there should be no barred galaxies out here,

153
00:15:38,110 --> 00:15:43,360
the disks of galaxies out beyond the redshift of what dynamically hot particularly

154
00:15:43,360 --> 00:15:48,910
in the gas we know that from work done with came on another Oxford type projects.

155
00:15:49,870 --> 00:15:53,230
And so those disks, it's perhaps surprising, maybe it's sustainable.

156
00:15:53,470 --> 00:16:02,140
But you also see this this change in morphology. We're beginning to be able to do serious comparison across a very large redshift,

157
00:16:02,140 --> 00:16:09,010
large enabled by the fact that our classification scheme is the same in both cases, because we have to take into account bias,

158
00:16:09,340 --> 00:16:13,540
we have to deal with the fact that these are distant galaxies and that faint, fuzzy,

159
00:16:13,540 --> 00:16:19,270
distant things tend to look featureless, but we can measure all of that correct for it and get these robust results.

160
00:16:20,290 --> 00:16:24,550
I just want to take a few minutes to I can't give you all the science highlights of galaxies.

161
00:16:24,970 --> 00:16:30,820
There are far too many, but I wanted to highlight one particular story that we've been following here in recent years.

162
00:16:32,620 --> 00:16:40,030
This is a nearby galaxy rejoices in the name NGC 4395 and it's a disk galaxy.

163
00:16:40,480 --> 00:16:48,040
You probably call it a spiral that what it is really is a flatulent galaxy, my favourite kind of galaxy, because it's fun to say the word flatulent.

164
00:16:48,430 --> 00:16:55,410
I recommend you all try it later. We can do it together now if you like to try.

165
00:16:55,540 --> 00:17:04,060
Try it in the privacy of your own homes later. But what you'll notice is that this thing is missing a classic component of a typical spiral galaxy.

166
00:17:05,440 --> 00:17:10,299
Patrick Moore is used to give public talks in which he described the Milky Way as true.

167
00:17:10,300 --> 00:17:14,260
Fried eggs clapped back to back an experiment I don't recommend.

168
00:17:14,560 --> 00:17:21,280
But the point being that you've got the disk and then you have this bulge of older stars at the centre and this is a truly bulge less galaxy.

169
00:17:22,510 --> 00:17:26,860
It's one of two things that are strange about it. One thing that's all there is that it's bulges.

170
00:17:27,160 --> 00:17:31,120
The second thing that's strange is that the black hole at its centre,

171
00:17:31,540 --> 00:17:35,379
it's going to call it a supermassive black hole, but it barely qualifies for that title.

172
00:17:35,380 --> 00:17:43,570
It's only about 300,000 solar masses, about a factor of ten lower than you'd expect for a galaxy of this mass.

173
00:17:45,160 --> 00:17:48,670
So could these things will be connected? Well, yes, there's a nice story here.

174
00:17:48,670 --> 00:17:54,070
So the fact that it's bold to us tells you that this galaxy is guaranteed merger free.

175
00:17:55,120 --> 00:18:01,390
So the simulators and the dynamics tell us that if you have any even minor merger,

176
00:18:01,480 --> 00:18:06,340
maybe something with a something a 10th of the mass of this galaxy comes in and collides with it,

177
00:18:06,640 --> 00:18:11,020
then that inevitably kicks stars out of the disk and into a bulge.

178
00:18:11,410 --> 00:18:20,440
And so a bulge is galaxy is a laboratory for what happens if you let a galaxy evolve in isolation instead of in the presence of multiple mergers.

179
00:18:21,220 --> 00:18:30,250
Now, one can speculate people have made good arguments that it's the merging of different galaxies that

180
00:18:30,250 --> 00:18:35,980
drives both the growth of the galaxy themselves and the growth of the black holes at the centre.

181
00:18:36,160 --> 00:18:41,860
And that this explains the tight relation that we see between the mass of a galaxy and the mass of it's black hole.

182
00:18:41,860 --> 00:18:47,320
So the fact that this is pulseless and has a low black hole fits that theory.

183
00:18:47,560 --> 00:18:53,920
Low mass black hole fits that theory rather well. But it's one galaxy amongst the galaxy do we could do better than that.

184
00:18:53,920 --> 00:18:57,070
We ask the question, how prominent is the central bulge?

185
00:18:57,460 --> 00:19:01,000
We could pick those four where people say there's no bulge or there's just noticeable.

186
00:19:01,690 --> 00:19:03,820
We actually then for this particular study,

187
00:19:04,690 --> 00:19:12,700
restrict ourselves only to those where we're utterly convinced there's no bulge and which have currently growing black holes.

188
00:19:13,180 --> 00:19:20,500
ADRIANNE So these are bulges, galaxies, merger free, which are currently active.

189
00:19:20,860 --> 00:19:25,030
And you can see that because they have point sources at the centre. So these aren't small bulges.

190
00:19:25,030 --> 00:19:32,920
If you do the modelling, these things are just the gas around the accreting black hole.

191
00:19:33,370 --> 00:19:40,299
Now we'll get rid of two of them because those are mergers about to start and then we can look at the properties of these galaxies.

192
00:19:40,300 --> 00:19:45,720
So here they are. This is the black hole mass and this is the bulge stellar mass.

193
00:19:45,730 --> 00:19:48,910
And we can directly measure the black hole mass for two of these systems.

194
00:19:49,870 --> 00:19:58,540
The rest we can put limits on by assuming that they're accreting at the Eddington luminosity, which is hard but not always about limit.

195
00:20:00,170 --> 00:20:02,210
That's a little unfair to these guys.

196
00:20:02,690 --> 00:20:09,950
And if you measure the accretion rate or the luminosity rather from these two and you assume that that's typical of the whole population,

197
00:20:10,490 --> 00:20:14,090
well, these things move and you get a picture that looks like this.

198
00:20:14,360 --> 00:20:18,950
And this is, I think, the least surprising graph I could possibly show.

199
00:20:19,160 --> 00:20:24,739
We've selected galaxies because they're bulge bulges and we find that for a given black hole mass.

200
00:20:24,740 --> 00:20:28,550
Sorry, this black stripe is normal galaxies. From Harring and Rex.

201
00:20:29,690 --> 00:20:35,930
We selected these galaxies because they're faultless. We find that for a given black hole mass, they have smaller bulges than they should do.

202
00:20:35,960 --> 00:20:40,250
This is not an exciting result. But if I plotted instead of the black hole mass.

203
00:20:40,970 --> 00:20:44,510
So it's the bulge mass, the total mass of the galaxy.

204
00:20:44,840 --> 00:20:52,340
These things look normal. So if you if the bulge is taken out, plan, all you care about is the evolution of the galaxy.

205
00:20:52,610 --> 00:21:01,370
It's the first galaxy that I showed that's the anomaly and that we can grow perfectly normal galaxies with normal sized black holes without mergers,

206
00:21:02,240 --> 00:21:09,050
at least in these cases. We have a huge programme of work to try and understand whether these are freaks and unusual in some way

207
00:21:09,380 --> 00:21:13,820
or whether this is a mode of galaxy and black hole growth that we need to care about more generally.

208
00:21:15,560 --> 00:21:21,710
But it starts with morphology. But let's return to the fundamental thing again.

209
00:21:22,000 --> 00:21:26,479
Galaxies. Those are details. You know, we talked about bars and bulges,

210
00:21:26,480 --> 00:21:31,459
but the fundamental morphological difference of the population of galaxies is are they spiral or elliptical?

211
00:21:31,460 --> 00:21:39,650
And all sorts of properties correlate with these two spirals tend to but all that exclusively be star forming which are lots of red spirals.

212
00:21:39,950 --> 00:21:46,100
Ellipticals tend to be old, red and dead, devoid of both the gas and of the new stars that might form from that gas.

213
00:21:46,400 --> 00:21:49,580
Although we have a whole load of blue ellipticals that I could tell you about as well.

214
00:21:50,360 --> 00:21:54,380
But splitting the population into these two systems helps you understand it.

215
00:21:55,190 --> 00:21:58,849
So in the top left here, this is the I like to think of.

216
00:21:58,850 --> 00:22:02,149
This is the Hertz from Russell Diagram of Galaxy Evolution.

217
00:22:02,150 --> 00:22:06,460
It's how we understand these things. So this is mass versus colour.

218
00:22:06,470 --> 00:22:13,880
So blue galaxies are down here, red up here, small galaxies, big galaxies, or if you prefer, less luminous and more luminous galaxies.

219
00:22:14,330 --> 00:22:17,629
And then in the controls you see the two features of this landscape.

220
00:22:17,630 --> 00:22:25,280
We see the blue cloud where normal galaxies spiral galaxies, and the red sequence composed primarily of ellipticals.

221
00:22:25,970 --> 00:22:31,450
It's kind of interesting that the Milky Way is in a strange place on this diagram where green and the green,

222
00:22:31,460 --> 00:22:37,640
but you can see and then what I've plotted in colour here in this slightly making green sorry about

223
00:22:37,640 --> 00:22:44,240
that is the fraction of galaxies at that point in the diagram but have actively growing black holes.

224
00:22:44,990 --> 00:22:52,730
And so you can see if you look to all galaxies and eventually you conclude it was the intermediate mass galaxies where the action was happening.

225
00:22:52,820 --> 00:22:57,140
But if we split into up in the top right elliptical galaxies,

226
00:22:57,230 --> 00:23:05,780
early types you see it's the least massive the in the present day have 18 and if you look at light type spirals you see it's the most massive spirals.

227
00:23:06,470 --> 00:23:12,410
But the picture you use you talk about is different depending on whether you're looking at ellipticals or spirals.

228
00:23:13,190 --> 00:23:15,020
And we could see this in other places as well.

229
00:23:15,260 --> 00:23:23,749
Kevin Quincey of the The Kevin Week wrote a paper last year with the Galaxy Z team and Data which rejoices in the title.

230
00:23:23,750 --> 00:23:31,970
The Green Valley is a red herring. Now there's literature that shows that silly titles don't get more citations,

231
00:23:32,360 --> 00:23:36,470
but this thing has been cited more than 30 times in its first six months, so it must be a very good paper.

232
00:23:36,830 --> 00:23:41,540
Um, and what you can see is a sort of equivalent of the diagram I just been showing instead of mass.

233
00:23:41,540 --> 00:23:47,210
This is colour, we've got colour here blue, red and this is now an ultraviolet colour.

234
00:23:47,690 --> 00:23:52,340
So this is brightest in the ultraviolet fights in the ultraviolet, brightest in the ultraviolet.

235
00:23:52,940 --> 00:23:57,260
And you see once again that we have this nice bimodal distribution.

236
00:23:58,040 --> 00:24:01,910
And the question that Kevin is trying to answer derives from the previous plot.

237
00:24:02,300 --> 00:24:05,720
He's trying to work out how galaxies cross the Green Valley,

238
00:24:05,930 --> 00:24:10,760
this region with relatively few galaxies that lives between the blue cloud of the red sequence.

239
00:24:11,870 --> 00:24:16,670
And in particular, he's interested in the question of how quickly galaxies could cross.

240
00:24:18,020 --> 00:24:22,370
So what he does is he sets up a toy model. He says, okay, these galaxies have constant star formation, right?

241
00:24:22,970 --> 00:24:32,000
And then something quenches star formation by some exponential process and the speed with which quenching can happen is left as a free parameter.

242
00:24:32,420 --> 00:24:36,350
And then then he thinks, okay, let's think about what that would do.

243
00:24:36,800 --> 00:24:42,560
If all your galaxies start blue and then you quench them, the big blue stars start dying off.

244
00:24:42,560 --> 00:24:48,500
You're not replacing them with star formation because of the quenching, and you find they follow these tracks and in a sensible amount of time.

245
00:24:48,920 --> 00:24:54,620
In the 4 billion years since he locates this quenching, you find that if you have a very rapid quench,

246
00:24:54,620 --> 00:24:59,270
they get all the way up here and they're red in optical colour and that red in the ultraviolet as well.

247
00:24:59,740 --> 00:25:06,910
Whereas if you have slow processing, they stay blue. Sometimes we can do good science with very simple models.

248
00:25:07,390 --> 00:25:12,820
And then he says, okay, let's look at where green spirals and green electricals are.

249
00:25:13,660 --> 00:25:19,930
You see that here? These are the green spirals. There are bands fit with this rapid quenching model.

250
00:25:19,960 --> 00:25:25,120
The rapid quenching model leaves you with galaxies that are too red rather than to the green.

251
00:25:25,690 --> 00:25:29,710
And so it's a poor fit for the spirals. So spirals must quench only slowly.

252
00:25:30,820 --> 00:25:36,040
The ellipticals in orange up here can be fit by those fast quenching models, but not by the slow ones.

253
00:25:36,040 --> 00:25:37,449
So ellipticals quench faster.

254
00:25:37,450 --> 00:25:45,310
So to keep in the Green Valley is a red herring because it's just the tails of these two distributions slow quenching spirals,

255
00:25:45,520 --> 00:25:48,670
rapidly quenching ellipticals. Another difference in morphology.

256
00:25:50,830 --> 00:25:54,340
It's a good result. It's an interesting result. It's inspired a lot of work.

257
00:25:54,610 --> 00:26:00,189
But slightly embarrassingly in this paper, we do this by, we look at this diagram, but we say,

258
00:26:00,190 --> 00:26:04,719
Yeah, that's a good fit and that's not really good enough anymore, so do it properly.

259
00:26:04,720 --> 00:26:08,770
You ask a PhD student to do it. Becky Smethurst, who works with me,

260
00:26:08,770 --> 00:26:18,090
has just fitted a full Bayesian MCM model of galaxy quenching to the entire galaxies population using incidentally that fuzzy logic.

261
00:26:18,100 --> 00:26:24,190
The fact that we know that this is a spiral galaxy, but only seven out of ten people said it was a spiral.

262
00:26:24,190 --> 00:26:33,009
So using that to quantify how strong each galaxies evidence is and this is a plot of the time at

263
00:26:33,010 --> 00:26:39,249
which we expect quenching from the present day back to the beginning or near to the Big Bang.

264
00:26:39,250 --> 00:26:42,340
We have a strike. We have strategy and physical models.

265
00:26:42,790 --> 00:26:45,969
But don't let that distract you and then the degree of quenching from slow.

266
00:26:45,970 --> 00:26:51,820
Patrick The first question and the point here is just to point out that for green galaxies, blue galaxies,

267
00:26:51,820 --> 00:26:58,479
red galaxies, for disk galaxies and for smooth galaxies, a wide variety of quenching rates exist.

268
00:26:58,480 --> 00:27:02,590
So both ellipticals and spirals have fast and rapid quenching.

269
00:27:03,160 --> 00:27:08,140
Which brings me to Brian May. Red Star, Resolute, Young quenched.

270
00:27:08,290 --> 00:27:12,520
Even after 30 years. I've only just thought of that. So glad. I thought it would get a laugh, but it hasn't.

271
00:27:12,610 --> 00:27:18,099
So the only thing I could do now is make this sort of a matter joke between me and you as the audience.

272
00:27:18,100 --> 00:27:20,710
So thank you for going with that. I'm Brian.

273
00:27:21,460 --> 00:27:31,240
As many of you know, was a Ph.D. student at Imperial College working on the zodiacal light on dust in the universe, dust in the solar system.

274
00:27:31,240 --> 00:27:36,190
Rather, when his rock career took off, he took a 30 year break, and at that point,

275
00:27:36,190 --> 00:27:41,770
Imperial offered him an honorary degree, which he turned down on the grounds that he wanted to finish his thesis.

276
00:27:43,690 --> 00:27:46,749
And I've been known to say at this point this has been recorded.

277
00:27:46,750 --> 00:27:51,579
So I am not saying that it helps if you want to take a 30 year break, if you work on something so boring,

278
00:27:51,580 --> 00:27:58,510
but it's still there when you come back 30 years later, actually it's some fad, new observational techniques.

279
00:27:58,510 --> 00:28:04,650
And come on, Brian was actually able to go back and do more observing and do a longitudinal study over 30 years of the behaviour,

280
00:28:04,670 --> 00:28:08,980
the darker light which no one else could have done because no one else can fund themselves for 30 years.

281
00:28:09,340 --> 00:28:20,649
And he's got more nature papers than me. But as he was getting back into this idea, he has look it up, I've got more citations for.

282
00:28:20,650 --> 00:28:24,340
That's fine. Anyway, it's true.

283
00:28:25,270 --> 00:28:30,150
I, I at the time he was getting back into this, I was writing a popular book with pride,

284
00:28:30,340 --> 00:28:39,010
and I taught him how to use Astro from the archive as the modern because he set off for the library on his first day back in, back in research.

285
00:28:39,370 --> 00:28:45,730
And so in return he promoted galaxies a creating a permanent bias in the musical taste of our classifiers.

286
00:28:46,060 --> 00:28:50,110
This is a Dutch schoolteacher called Honey Barnacle, who likes bronze music.

287
00:28:50,110 --> 00:28:58,330
She's got the same guitar who discovered this thing. She's become a poster child for the other real advantage of having humans look at your data.

288
00:28:58,630 --> 00:29:04,990
So we've dealt with the fact that sometimes you need to scale to millions of people to cope with the size of modern data sets.

289
00:29:05,530 --> 00:29:10,000
But humans are also distractible, wonderfully distractible.

290
00:29:10,280 --> 00:29:16,600
We have this almost unique ability to be doing a routine task and then be brought up short by the unusual and unexpected.

291
00:29:17,080 --> 00:29:26,080
And this is surprisingly hard to program. You could build anomaly detection routines that get distracted mostly by common anomalies and Galaxy Z.

292
00:29:26,290 --> 00:29:28,150
You find an awful lot of satellite trails,

293
00:29:28,570 --> 00:29:36,190
but to build a classifier capable of doing a good job of classifying galaxies and then stopping and saying not just that this is a spiral galaxy,

294
00:29:36,430 --> 00:29:44,739
but that it's a blue blob here is rather hard. This blue blob, a quite the name of the that we thought it was a technical term.

295
00:29:44,740 --> 00:29:50,229
It turns out to be Dutch for object or thingy, but monthly notices let us print it.

296
00:29:50,230 --> 00:29:51,970
So this is now the official name of this object.

297
00:29:52,600 --> 00:29:59,130
And this was a return to what I thought astronomers did when I was playing around with my small telescope because we found.

298
00:29:59,200 --> 00:30:02,230
And a weird thing. And then we pointed lots of telescopes at it.

299
00:30:02,590 --> 00:30:06,250
And when we did that, we discovered, first of all, that it's at the same redshift as this galaxy.

300
00:30:06,760 --> 00:30:10,630
So it's distant and it's large. We discovered that it's hot.

301
00:30:11,350 --> 00:30:14,650
It's about 50,000 Kelvin. The gas in the vulva.

302
00:30:14,920 --> 00:30:18,309
And it contains no stars, or at least no obi stars.

303
00:30:18,310 --> 00:30:28,950
No bright stars. And so we suspected that this was being heated by a jet, it coming from activity at the nucleus of this galaxy.

304
00:30:28,960 --> 00:30:35,620
So you do see activity in spiral galaxies and we do sometimes see spectacular jets.

305
00:30:35,650 --> 00:30:43,090
Here's a discovery from the end of last year made by another Galaxies user on a version of the project that's looking at radio data.

306
00:30:44,020 --> 00:30:50,890
So the idea was that there would be jets of material that comes out here this cold see either side of the border.

307
00:30:50,920 --> 00:30:57,630
We know that from radio and that this is excited by the impact of the jets on the spectrum of the ball.

308
00:30:57,650 --> 00:31:01,300
VEP is consistent with it being shock heated, which would fit.

309
00:31:02,440 --> 00:31:09,490
The only problem is that when you look at this galaxy with an X-ray telescope, you discover that there's almost no source at all.

310
00:31:09,820 --> 00:31:14,680
So there isn't a luminous enough source, even with ridiculous obscuration by dust,

311
00:31:15,070 --> 00:31:19,390
to have an active and powerful enough heating of this amount of gas right now.

312
00:31:20,390 --> 00:31:26,590
And so the conclusion that we came to was that this the vulva was a light echo of this.

313
00:31:26,620 --> 00:31:34,000
It was telling us about the state of the galaxy 50,000 years ago, because there's a dogleg in here of about 50,000 light years.

314
00:31:34,510 --> 00:31:38,499
And so, in other words, 50,000 light years, 50,000 years ago, if we looked at the system,

315
00:31:38,500 --> 00:31:42,210
we would see a bright again would be magnitude eight in the visible.

316
00:31:42,220 --> 00:31:48,370
You'd be able to see it in binoculars, and it would, in fact be the nearest quasar to the Milky Way.

317
00:31:48,880 --> 00:31:53,560
But in the last 50,000 years, it's shut down and it's gone from active to quiet.

318
00:31:53,560 --> 00:31:58,870
Something that we knew happened, but which was very difficult to catch in the act.

319
00:31:59,770 --> 00:32:01,780
And so this is great fun. It's an interesting story.

320
00:32:01,780 --> 00:32:08,019
We've got a new way of studying the shut down of an AGM because we can read off the history of about 50,000

321
00:32:08,020 --> 00:32:13,960
years worth of the radiation from this black hole by looking at different distances along the vulvar.

322
00:32:16,520 --> 00:32:21,770
But the explanation smelled funny because why would the nearest quasar just shut down?

323
00:32:22,760 --> 00:32:28,400
Unless this happens all the time. And so these are the four values, which is the diminutive of a vulva.

324
00:32:28,730 --> 00:32:30,410
It's amazing what you learn when doing this stuff.

325
00:32:30,650 --> 00:32:38,670
And these are all the blue things you can see in these images are 18 ionised clouds in a variety of slowing galaxies, some a powerful AGM.

326
00:32:39,290 --> 00:32:47,180
But a third of them are not. And so we think a third of these systems have shut down within the last 20 to 100000 years.

327
00:32:47,540 --> 00:32:52,429
And we have a Hubble Space Telescope program to look at them and to try and work out

328
00:32:52,430 --> 00:32:57,290
the geometry of what is actually quite a strange and fascinating set of objects.

329
00:32:57,290 --> 00:33:01,790
The complexity of these images is both astounding and deeply annoying,

330
00:33:02,090 --> 00:33:05,540
because if you've got a really simple geometry, you can read off the history much more easily.

331
00:33:06,050 --> 00:33:13,930
But in step, we've got this mess now. So hopefully by this point, I've convinced you that I think galaxies are interesting.

332
00:33:13,940 --> 00:33:22,579
That's the subtext. But also that getting people involved in the sort of crowdsourcing and what we call citizen science has benefits,

333
00:33:22,580 --> 00:33:27,020
not just because of scale, but because of the potential for serendipity as well.

334
00:33:29,000 --> 00:33:33,380
So the next thing you might wonder is how else you can apply this. This is obviously an interesting method.

335
00:33:33,890 --> 00:33:38,120
It was also free and it also has huge outreach potential.

336
00:33:38,120 --> 00:33:41,390
So what else can we do? First question, Oh, are we going to run out of people?

337
00:33:41,840 --> 00:33:46,430
Is that a serious concern that we should worry about? And then my argument is that it isn't.

338
00:33:47,120 --> 00:33:48,829
Every time you talk about an Internet project,

339
00:33:48,830 --> 00:33:55,310
you have to mention a guy called Clay Shirky who writes annoyingly bestselling books about how the Internet is changing everything.

340
00:33:55,550 --> 00:34:05,180
This is my Clay Shirky slide. The big box shows the 200000000000 hours American adults spent watching television in 2009.

341
00:34:05,600 --> 00:34:09,200
And the small box is 100000000000 hours it took to create Wikipedia.

342
00:34:10,430 --> 00:34:15,650
And that, if you want to be facetious about it, Angry Birds use 16 years of human attention every hour.

343
00:34:16,400 --> 00:34:21,620
It would be much better for science if Angry Birds was the study of avian behaviour than just a game.

344
00:34:21,620 --> 00:34:33,799
So we only need a small, tiny fraction of the attention that's out there to be used for good, to get huge processing power so we can scale this.

345
00:34:33,800 --> 00:34:39,440
We can, if we want to, without harming the progress of astrophysics, have people look at penguins.

346
00:34:40,520 --> 00:34:47,299
So one of the things that happened to me once we started doing this was the other researchers would sidle up to me or call or email and say,

347
00:34:47,300 --> 00:34:52,340
You know, your people do like looking at galaxies. Do you think they'd look at my data as well?

348
00:34:53,120 --> 00:35:00,830
And in most cases, the answer is yes, because the people who looked at galaxies were motivated by the desire to contribute to science.

349
00:35:01,070 --> 00:35:06,260
And so this is an image from a project run out of the Department of Zoology here in Oxford.

350
00:35:06,620 --> 00:35:13,759
They used to visit remote parts of the Antarctic by yacht, and I'm told the yacht is not as luxurious as that makes it sound.

351
00:35:13,760 --> 00:35:16,940
In the same way that when I say I'm going to Hawaii, it doesn't mean I'm going to the beach.

352
00:35:18,920 --> 00:35:23,130
Sometimes I go to the beach. They used to visit Count penguins.

353
00:35:23,180 --> 00:35:27,020
Penguin colonies are sensitive to climate change, so this is interesting.

354
00:35:27,020 --> 00:35:29,270
And they revisit the same colony every year or two.

355
00:35:30,320 --> 00:35:37,580
Now they can buy a cheap camera for 100 quid, leave it out, and it will take pictures every 5 minutes or 5 hours.

356
00:35:37,580 --> 00:35:39,950
And these cameras survive the Antarctic winter.

357
00:35:40,250 --> 00:35:47,150
So now they still got one Ph.D. student, but instead of a notebook filled with penguin counts, they come back with hard drives full of images.

358
00:35:47,900 --> 00:35:52,309
So the task here is to count the number of Penguins game machine learning.

359
00:35:52,310 --> 00:35:56,330
Can't do a very good job of this. Anyone want to tell me what the answer is?

360
00:35:58,450 --> 00:36:01,720
Three three. It depends whether you count integer penguins or not.

361
00:36:01,840 --> 00:36:07,060
I agree. I agree that three is an upper limit, but you'll be amazed how often people check out decimals.

362
00:36:07,660 --> 00:36:11,080
All right, so you're warmed up. So the next one is a speed test.

363
00:36:11,590 --> 00:36:15,160
So I want the first person to count the penguins in the next image. Just call out the answer.

364
00:36:19,700 --> 00:36:21,530
I'm just going to the pub while you work that.

365
00:36:21,740 --> 00:36:29,150
So the point is, this is one of the interesting things about this task is that there's a wide variety of difficulties.

366
00:36:29,390 --> 00:36:37,610
This is a more normal image. And here's the solution. We built a project called Penguin Watch, which lets you count penguins in your spare time.

367
00:36:39,320 --> 00:36:44,030
Surprisingly enough, this has proved rather popular. More than a million penguins have been counted,

368
00:36:44,420 --> 00:36:50,660
and I feel the need to point out that we considered but rejected the name Penguin Hunters for this project,

369
00:36:51,470 --> 00:36:57,080
although I think that would have been still more popular. And we can go from penguins to particle physics.

370
00:36:58,010 --> 00:37:01,040
Not. And obviously it's not funny. Penguins are funny. Particle physics?

371
00:37:01,040 --> 00:37:05,320
No. Yeah. Maybe I need to. Subterranean physics jokes.

372
00:37:06,140 --> 00:37:09,350
This is a project called Higgs. Or this is data from Atlas. Right.

373
00:37:09,350 --> 00:37:12,620
But this is data from Atlas. From a very special source.

374
00:37:12,890 --> 00:37:20,090
So you're seeing a cross-section through the LHC here, of course, a collision and a particle cascade coming from a collision.

375
00:37:20,510 --> 00:37:27,530
Most of the LHC data doesn't hit a hard drive, as I'm sure lots of you know it's safe if it satisfies a series of triggers.

376
00:37:28,220 --> 00:37:36,230
And in the the debt detritus in the waste that doesn't hit those triggers is mostly a lot of routine stuff, some nonsense.

377
00:37:37,010 --> 00:37:46,610
I'm just possibly a few Nobel Prize winning discoveries because anything really unexpected would end up missing the triggers as well.

378
00:37:46,650 --> 00:37:50,600
I think this is a project this is almost the first project that Joe proposed we build.

379
00:37:51,200 --> 00:37:55,610
When I came to talk to him about Galaxy Z, so we finally did it, which is good.

380
00:37:56,450 --> 00:38:00,860
And the point about this is this is here for two reasons. One is it's kind of interesting.

381
00:38:01,640 --> 00:38:07,520
It really gets the sense of what the Atlas data is like. But the other thing is it's a test of how well we understand our community.

382
00:38:08,030 --> 00:38:13,700
The many also volunteers we have on our projects have told us that they do these projects because they want to help science.

383
00:38:16,030 --> 00:38:22,770
This is an incredibly long shot. It's scientifically useful, but this project will almost certainly fact it succeeds.

384
00:38:22,780 --> 00:38:28,690
Then there will be Nobel Prizes. And on the first day that we launched, I got this email from Alan.

385
00:38:29,290 --> 00:38:34,780
Look, if Alan's here. I saw him early. I think he's escaped. And so this was about this image that you just say, Who's messing with us?

386
00:38:35,110 --> 00:38:38,889
We've got a jet of new ones, which is a feature predicted in sum beyond the standard model theories.

387
00:38:38,890 --> 00:38:42,879
But it's never been seen in the wild. I don't remember us put together things like this into the simulation.

388
00:38:42,880 --> 00:38:46,209
Is this real? So I thought in an hour we've.

389
00:38:46,210 --> 00:38:49,830
We've solved this problem. It turns out this is actually just a breakthrough.

390
00:38:49,840 --> 00:38:52,540
These are normal particles that have escaped. It's not a jet of neurones.

391
00:38:53,260 --> 00:38:57,760
We're still looking for Nobel Prizes, but the project is quite popular and I recommend you give it a go.

392
00:38:57,970 --> 00:39:01,140
And you've got the idea by now that we can try this on all sorts of things.

393
00:39:01,150 --> 00:39:04,209
Probably the furthest distance we've travelled from galaxies.

394
00:39:04,210 --> 00:39:09,370
There is a project led by Brooks Simmons. It will also let the pulses galaxy work I talked about.

395
00:39:09,940 --> 00:39:11,769
But this is the Planetary Response Network.

396
00:39:11,770 --> 00:39:23,800
We have a partnership with providers of satellite imagery to react to a disaster like a typhoon hitting a developed area

397
00:39:24,100 --> 00:39:32,050
to provide rapid assessment of the damage and destruction to teams on the ground who need this kind of information.

398
00:39:32,350 --> 00:39:35,820
So we're testing that now if you want to join in. There's a go around.

399
00:39:35,830 --> 00:39:39,580
If you go and test, we're testing it with the historical data because clearly in this circumstance,

400
00:39:39,850 --> 00:39:44,200
you really need to demonstrate that you've got your data analysis right before you deploy the system.

401
00:39:44,920 --> 00:39:49,210
All of these things live in an umbrella site called the Zooniverse.

402
00:39:50,020 --> 00:39:56,950
Word I hope you've heard. I was reminded by Rob Simpson the other day that we have 42 Zooniverse projects have launched.

403
00:39:56,950 --> 00:40:01,780
They all lots of them have dice with the logos can see there's a sort of acceleration and then a fairly steady pace.

404
00:40:02,500 --> 00:40:09,850
This is felt very rapid and you can also detect in this graph and times where we've made changes.

405
00:40:09,850 --> 00:40:17,350
So for example, this cluster of projects were built using software or built in-house that we called Juggernaut.

406
00:40:17,650 --> 00:40:24,280
And Juggernaut was a piece of software that ran one project. So you built a Galaxy Zoo juggernaut and then you use much of the same code,

407
00:40:24,280 --> 00:40:27,910
but you built a lunar juggernaut and then so on and so upon it, hunting juggernaut.

408
00:40:28,360 --> 00:40:33,669
And it turns out that by the time you get to this many projects, you spend all of your time maintaining those pieces of code.

409
00:40:33,670 --> 00:40:36,700
So you throw that out and we build code called a robot us.

410
00:40:37,570 --> 00:40:38,110
Don't worry about that.

411
00:40:38,590 --> 00:40:47,890
And The Robots is one app that runs most of our projects that work well, but we were getting stuck at the limit of about 20 projects.

412
00:40:48,280 --> 00:40:58,149
And so the difference between penguins and planet and penguins and particle physics is small enough that you can further generalise the problem.

413
00:40:58,150 --> 00:41:06,550
So we have a new code called Panopto is Panopto is is a whole set of services that will allow people to build and run their own projects.

414
00:41:07,090 --> 00:41:14,710
So instead of coming to us and having us develop a new project, what we'll do instead is for projects where the interface is simple.

415
00:41:15,250 --> 00:41:21,790
You'll just come along as a scientist, upload your data, choose from some options about what questions you want to ask,

416
00:41:21,790 --> 00:41:23,530
what tasks you want to sign, people to click a button to.

417
00:41:23,610 --> 00:41:28,899
You have your project live on the web where a month or two or three I'm sorry to see which developers are in the room.

418
00:41:28,900 --> 00:41:34,210
They're not here. So we're only about a month away from being able to put this life.

419
00:41:34,510 --> 00:41:42,819
If you have lots of data in the form of images or video or perhaps even sound that it would benefit you to have people look at Please,

420
00:41:42,820 --> 00:41:48,399
please come and talk to us. We need test cases for this stuff and we hope that the Zooniverse will go to hundreds or

421
00:41:48,400 --> 00:41:55,060
thousands of projects because we don't have to worry about yet about running out of people,

422
00:41:55,870 --> 00:41:59,949
because the more projects we do, we could treat each one as a natural experiment.

423
00:41:59,950 --> 00:42:02,559
And we've done easy projects of difficult projects.

424
00:42:02,560 --> 00:42:07,030
We've done projects where it takes 2 seconds to look at an image and projects where it takes 10 minutes.

425
00:42:07,360 --> 00:42:12,760
We've done projects where the images are beautiful and we've done projects where people look at graphs for fun.

426
00:42:14,340 --> 00:42:18,819
Fascinated me. The beauty of the images doesn't seem to matter, it just kind of reassuring,

427
00:42:18,820 --> 00:42:22,720
but backs up this idea that people are doing this because they want to help science.

428
00:42:23,350 --> 00:42:29,620
We also see general patterns emerging. So this is a graph showing the contributors to one particular project.

429
00:42:29,620 --> 00:42:34,629
I think this is one of the incarnations of galaxies in which box is a person.

430
00:42:34,630 --> 00:42:37,450
The colours are meaningless other for the purposes of making it look nice.

431
00:42:37,810 --> 00:42:44,230
And then you can see we get a huge number of classifications from these people over here on the left, each of whom does a huge amount of work.

432
00:42:45,220 --> 00:42:51,550
But we also get a lot of classifications for people who only breeze by is common to almost all of our projects.

433
00:42:51,820 --> 00:42:55,600
And we deliberately design so that both of these sets of people are useful.

434
00:42:55,930 --> 00:43:00,490
If you rely just on the experts, you never create enough people to become experts.

435
00:43:00,820 --> 00:43:04,090
You rely on these people. You miss out on a lot of the stuff.

436
00:43:04,600 --> 00:43:12,100
And so we now have this very harmonious situation in which machines collect the data and

437
00:43:12,100 --> 00:43:15,270
we have enough humans with the help of our citizen scientists to work through them.

438
00:43:15,680 --> 00:43:18,680
Machine and human always worked together really well.

439
00:43:21,730 --> 00:43:28,180
Sorry. Wrong picture. This is my version of that nightmare, because this harmonious situation isn't going to last.

440
00:43:28,240 --> 00:43:40,270
This is the large synoptic survey telescope. LCT, of which Oxford is an important member, is as big as the biggest telescopes today or less.

441
00:43:40,720 --> 00:43:46,870
But it's a survey telescope. So the scientists we were arguing about how it will move, but it'll scan the whole sky roughly every three nights.

442
00:43:47,230 --> 00:43:51,310
It will produce something like three terabytes of data a night.

443
00:43:52,090 --> 00:44:00,459
And another way of understanding that is to subscribe each of you to the closest, the alert service, which will send you a message.

444
00:44:00,460 --> 00:44:07,180
Whenever anything in the closest field changes, you'll get about two and a half text message, two and a half million text messages a night.

445
00:44:08,710 --> 00:44:12,880
There's an awful lot of stuff there and it's real. This is the top of a mountain being flattened.

446
00:44:13,120 --> 00:44:16,230
Flattened for Allah says to turn up a mirror exists.

447
00:44:16,240 --> 00:44:27,940
This project will start commissioning in 2018. Now if you want to study the common and the predictable, this isn't too much of a problem.

448
00:44:27,940 --> 00:44:35,049
So if you want to study, for example, conventional type want a supernovae somebody, some grad students and postdocs,

449
00:44:35,050 --> 00:44:39,129
somewhere some team of people will build a service that looks through those

450
00:44:39,130 --> 00:44:42,940
alerts and provides you with more type one supernovae than you can possibly use.

451
00:44:44,200 --> 00:44:45,910
We'll throw some out too, but that's okay.

452
00:44:46,120 --> 00:44:54,849
But if you want to understand the unusual, you want to be able to react in real time to truly, truly unpredicted things.

453
00:44:54,850 --> 00:45:00,670
We're going to need to keep people in that loop of classification left to use computers too.

454
00:45:00,850 --> 00:45:03,219
We have to get better, more efficient at using the people.

455
00:45:03,220 --> 00:45:06,700
So that's what we've been working on and that's what I want to talk about in my last few minutes.

456
00:45:07,360 --> 00:45:10,390
As an example, I'm going to use a project called Space Warps.

457
00:45:10,810 --> 00:45:17,200
PR is the three science prize apogee to Verma one of them who who's here in this department?

458
00:45:17,200 --> 00:45:19,599
Phil Marshall, who many of you will know spacewalks,

459
00:45:19,600 --> 00:45:26,680
is a search for gravitational lenses for these distant galaxies whose light has been bent by passage in front of a nearby galaxy.

460
00:45:26,920 --> 00:45:31,210
And the nice thing about this is it's really easy to calculate what these things should look like.

461
00:45:31,540 --> 00:45:36,760
And so we could put fake, but let's call them simulated galaxies into the system.

462
00:45:37,150 --> 00:45:40,150
And if you see one of these and you catch it, then we say, yes, well done.

463
00:45:40,720 --> 00:45:45,670
You call a simulated lens and you get some feedback and you gain confidence in your classifications.

464
00:45:46,780 --> 00:45:51,400
And then to really test the system, we found a use for particle physicist.

465
00:45:51,790 --> 00:45:57,850
So sorry. This is Brian Cox standing in front of the level telescope at Jodrell.

466
00:45:58,210 --> 00:46:02,620
We partnered with his program Stargazing Live to promote this project.

467
00:46:02,830 --> 00:46:07,450
And this is what happened when Brian Cox told people to go and look at space walks.

468
00:46:08,020 --> 00:46:13,940
This is a million constellations arriving within an hour, so it's a million images that have been sorted through.

469
00:46:14,710 --> 00:46:17,860
We can really do this at scale. And for each of these people,

470
00:46:18,250 --> 00:46:25,569
we get a sense of confidence about them because we see how they behave with these simulated images as on the x axis here,

471
00:46:25,570 --> 00:46:27,879
you've got the probability that somebody says there's a lens,

472
00:46:27,880 --> 00:46:32,740
that if there is a lens and on the Y axis you've got the probability that there isn't a lens.

473
00:46:33,280 --> 00:46:34,840
They say there isn't a lens when there is a one.

474
00:46:35,320 --> 00:46:40,209
So you can see most people up here on the top right do this is 50 randomly selected people in the top.

475
00:46:40,210 --> 00:46:44,320
Right. People do quite well. Our astute users are here in the bottom left.

476
00:46:44,320 --> 00:46:50,050
We've got some people doing less well, although note that if you know that somebody is wrong all the time,

477
00:46:50,590 --> 00:46:53,800
they're precisely as useful as somebody who's right all the time.

478
00:46:56,650 --> 00:47:05,200
A principle, I'm told, is understood by senior members of this department, although they won't give me any examples as to how they apply this.

479
00:47:05,380 --> 00:47:10,510
But you can also play different games. Everyone on the right of this diagram is an optimist, right?

480
00:47:10,510 --> 00:47:13,840
They tend to say that that's something that everyone on the left is a pessimist.

481
00:47:14,290 --> 00:47:18,909
And so you can start to think about how you might pass tasks around.

482
00:47:18,910 --> 00:47:24,760
If I've got an image that's been seen by a few people and I've begun to be confident that there's something there,

483
00:47:25,180 --> 00:47:31,660
I might pass that deliberately to a pessimist. And if they say that's something bad, but my confidence that it's real goes up dramatically.

484
00:47:31,930 --> 00:47:42,190
And one can quantify this and one can start simulating different modes of task distribution to optimise the number of classifications that you need.

485
00:47:42,850 --> 00:47:50,229
And there's definitely latency there. This is a complicated diagram from the space walks paper, but let's zoom in on the x axis.

486
00:47:50,230 --> 00:47:54,250
Just we're on this box. Sorry I didn't have time to make a new on the axis.

487
00:47:54,250 --> 00:47:57,700
This is the amount of effort it looks like. So these are people who do ten galaxies.

488
00:47:57,700 --> 00:48:02,229
100,000, 10,000, 100,000. And this is the skill.

489
00:48:02,230 --> 00:48:07,420
And the skill is the integrated information contributed.

490
00:48:07,420 --> 00:48:10,450
So we can actually calculate the channel information for each classification.

491
00:48:10,870 --> 00:48:16,810
And then we say, okay, the scale is the integrated amount of information provided by each classifier.

492
00:48:17,140 --> 00:48:20,290
And the first thing you notice is that we've done a good job in building this.

493
00:48:20,320 --> 00:48:23,350
Project. There's no one in the bottom, right?

494
00:48:23,680 --> 00:48:27,130
So no one is doing tens of thousands of notifications badly.

495
00:48:28,750 --> 00:48:30,640
So our feedback and our training works.

496
00:48:30,910 --> 00:48:38,620
But also note that there are all these people on the top left who only do a few classifications but who are very good at the minute.

497
00:48:38,620 --> 00:48:47,770
They see a random image, but if we could detect them, we could spend the little time we have with them on the things where we really need expertise.

498
00:48:48,190 --> 00:48:53,590
And so this is a sign that the project is not yet operating at full efficiency and we could take advantage of that.

499
00:48:54,370 --> 00:49:00,070
What's nice is that once you start doing that, once I start deciding who should see what,

500
00:49:00,640 --> 00:49:05,440
there is no reason that I've just built a system that's perfect for combining

501
00:49:05,440 --> 00:49:12,010
human and machine classifiers because I can take the behaviour of a machine.

502
00:49:12,940 --> 00:49:15,489
What fraction does the machine get right is a function of brightness,

503
00:49:15,490 --> 00:49:22,750
of image or something and decide whether to pass a particular image to a machine or to a human, the two on an equal footing.

504
00:49:23,560 --> 00:49:27,910
And we can cope with wild fluctuations in the number of people who are around.

505
00:49:28,210 --> 00:49:33,840
And so this is the system I think we need to build. Fred Astaire And this is what we're working through.

506
00:49:34,120 --> 00:49:37,180
But there are problems with both halves of this equation.

507
00:49:37,810 --> 00:49:42,850
There has been an enormous advance in machine learning in the last ten years.

508
00:49:43,240 --> 00:49:48,520
You know this because voice recognition basically works on your iPhone, right?

509
00:49:49,270 --> 00:49:54,160
Or systems that recognise your scribbled handwriting onto a tablet work pretty well.

510
00:49:55,750 --> 00:50:00,210
This is a schematic diagram that I don't want you to read really, but is that, you know,

511
00:50:00,400 --> 00:50:08,260
most of this work has been done through developments based on old style neural nets that involve what's called deep learning.

512
00:50:08,980 --> 00:50:16,180
So it's sort of a nested neural network procedure that does sophisticated things in understanding what it does and doesn't know.

513
00:50:17,350 --> 00:50:24,280
The universal problem with these methods is that they rely on extremely large training sets.

514
00:50:25,300 --> 00:50:33,070
If you don't have a large training set, you pass, say, 100,000 examples across five galaxies to one of these routines.

515
00:50:33,280 --> 00:50:40,980
What it does is it learns the individual characteristics of those 100,000 and then fails on anything that isn't exactly like those things.

516
00:50:41,440 --> 00:50:51,360
We ran a competition with 10,000 quid or dollars in prize money for machine learning, people to enter galaxies, to sort through galaxies.

517
00:50:51,370 --> 00:50:59,740
They would try and improve automatic classification of galaxies. And we heard back in a great uniform how the 200,000 example galaxies was not enough,

518
00:51:00,370 --> 00:51:04,660
and the person that won the competition did so by citing new data.

519
00:51:05,230 --> 00:51:12,250
So he realised that the problem with Galaxy classification is rotation, the invariant and rotated all of the images four times.

520
00:51:12,250 --> 00:51:15,250
So you got four different images and then chop them up in different ways.

521
00:51:15,250 --> 00:51:19,150
So from a single image he got 16 expert classifications.

522
00:51:19,420 --> 00:51:22,899
It's very clever actually and was able to try this data on that.

523
00:51:22,900 --> 00:51:29,500
So machines I think have this limit and astronomical surveys, particularly for rare objects,

524
00:51:29,740 --> 00:51:34,510
are not going to produce large enough training sets to allow us to take advantage of modern machine learning.

525
00:51:35,290 --> 00:51:40,749
So the idea that all of these problems are going to go away because the computer scientists have our backs is,

526
00:51:40,750 --> 00:51:45,580
I think, false for common stuff, maybe for simple stuff, maybe.

527
00:51:45,880 --> 00:51:49,960
But for the kind of stuff we're interested in, I don't think we can rely on them to solve the problem entirely.

528
00:51:50,170 --> 00:51:53,830
They're going to do the easy stuff and the rest of it will have to be sorted by people.

529
00:51:54,100 --> 00:51:57,220
But there's a problem with this optimisation for people as well.

530
00:51:57,760 --> 00:52:03,219
This is the performance of a classifier. I just look out on a supernova project, look at the blue light.

531
00:52:03,220 --> 00:52:06,430
Good performance puts this blue line near this top vertex here.

532
00:52:06,790 --> 00:52:09,220
So close to this dot is to the top and this is over time.

533
00:52:09,820 --> 00:52:16,870
And so this is a person who's been well trained and then they change their mind about something and they suddenly become a much worse classifier.

534
00:52:17,770 --> 00:52:21,009
And all my knowledge about this person is now, oh, only go back as well.

535
00:52:21,010 --> 00:52:24,940
So who knows what happened. Maybe they had a drink, maybe they handed their log in to somebody else.

536
00:52:25,210 --> 00:52:27,070
Maybe they were confused about the astrophysics.

537
00:52:27,730 --> 00:52:33,129
So there's very little understanding of how to cope with this because machines don't do this by train.

538
00:52:33,130 --> 00:52:36,400
A machine learning classifier does not suddenly change its mind.

539
00:52:36,760 --> 00:52:41,110
And so how do we detect that? How much effort do we put into detecting that sort of behaviour?

540
00:52:41,680 --> 00:52:45,990
The other problem with humans is that they have opinions about what they like to see.

541
00:52:46,630 --> 00:52:49,959
So the project called the Milky Way Project, which Rob Simpson led,

542
00:52:49,960 --> 00:52:55,420
looking for bubbles around newly formed stars in infrared surveys of our galaxy from Spitzer.

543
00:52:56,200 --> 00:53:03,010
These four images are the Milky Way Project images which most encouraged people to keep classifying.

544
00:53:03,340 --> 00:53:07,210
So we looked at the average number of tasks completed by somebody after seeing an image.

545
00:53:07,540 --> 00:53:14,889
So these things people found interesting and exciting and they're happy to classify and these are the worst behaviours.

546
00:53:14,890 --> 00:53:18,430
They're are a mixture of the confusing at the top and the boring of the bottom.

547
00:53:18,760 --> 00:53:27,500
So I have to be careful if. These are the most difficult pictures in the Milky Way project and I assign them to my best classifiers.

548
00:53:28,400 --> 00:53:28,969
Who are that?

549
00:53:28,970 --> 00:53:35,000
Partly because they even if they tell me that because they want to do science, but they're there because they want to see the pretty pictures.

550
00:53:35,210 --> 00:53:44,540
I've built a system that will systematically drive away my best classifiers and understanding how to automatically balance the requirements

551
00:53:44,540 --> 00:53:54,320
of people who want to be entertained and to feel confident versus just blindly optimising for efficiency is a huge and unexplored problem.

552
00:53:54,710 --> 00:53:59,600
And it's made worse because you can't ask people what they want to do.

553
00:54:00,290 --> 00:54:07,280
We have a problem with our community and it's a problem shared by anyone who tries to do science outreach.

554
00:54:07,290 --> 00:54:11,930
We think that this is a good way of getting people involved in science and thinking about science.

555
00:54:12,230 --> 00:54:16,370
But the people and we know that our volunteers take part because they want to help.

556
00:54:17,240 --> 00:54:21,860
But science people, you're actually doing science is a terrifying and scary thing.

557
00:54:22,520 --> 00:54:24,319
And so somehow we have to build projects.

558
00:54:24,320 --> 00:54:31,070
Does it feel real to convince people that they're doing something that will help astronomers without saying to them,

559
00:54:32,330 --> 00:54:40,940
the weight of the universe is on your shoulders? It's an incredibly difficult problem and it's taught me that galaxies are simply the people.

560
00:54:41,090 --> 00:54:41,720
Thank you very much.