1
00:00:00,660 --> 00:00:03,960
Module 10.2. 
Some research projects that use the

2
00:00:03,960 --> 00:00:07,770
techniques you have learned in the 
digital signal processing class.

3
00:00:09,050 --> 00:00:11,890
We're going to talk about some current 
research in the lab.

4
00:00:11,890 --> 00:00:15,982
There is a whole slew of them, and it's a 
selection here of interesting research

5
00:00:15,982 --> 00:00:21,642
projects that we can briefly discuss. 
The first one is, eFacsimile is a project

6
00:00:21,642 --> 00:00:26,670
on art work acquisition. 
The second one is about signal processing

7
00:00:26,670 --> 00:00:30,170
in sensor networks. 
Then there is a network science result on

8
00:00:30,170 --> 00:00:35,020
source localization put in graphs. 
Then we talk about sampling result, so

9
00:00:35,020 --> 00:00:38,690
called finite rate of innovation 
sampling.

10
00:00:38,690 --> 00:00:42,472
Then we talk again about sampling, that 
of physical fields, using some new

11
00:00:42,472 --> 00:00:47,260
techniques for sampling. 
Then, we have a project on image

12
00:00:47,260 --> 00:00:52,720
acquisition where we change the sensors 
used in acquiring images.

13
00:00:52,720 --> 00:00:55,270
Then, an old classic, predicting the 
stock market.

14
00:00:56,550 --> 00:01:00,587
Then, we talk about inverse problems. 
The three next projects are actually

15
00:01:00,587 --> 00:01:04,209
inverse problems. 
The first one is on the diffusion

16
00:01:04,209 --> 00:01:08,363
equation, the second one is trying to 
understand the nuclear fall out from

17
00:01:08,363 --> 00:01:14,400
Fukushima, and last but not least is an 
inverse problem in acoustics.

18
00:01:17,460 --> 00:01:20,480
The eFacsimile project. 
This is a project that we do together

19
00:01:20,480 --> 00:01:26,800
with Google to try to improve how artwork 
is represented on the internet.

20
00:01:26,800 --> 00:01:31,910
it's lead by [INAUDIBLE] researcher Loic 
Baboulaz and several PhD students are

21
00:01:31,910 --> 00:01:36,117
involved in this. 
The questions are how to capture,

22
00:01:36,117 --> 00:01:40,670
represent, and render artwork as well as 
possible.

23
00:01:40,670 --> 00:01:45,415
And to do this, we need some advanced 
techniques on relighting, manipulation

24
00:01:45,415 --> 00:01:50,452
of, the so called, light fields that is 
acquired, and potentially high resolution

25
00:01:50,452 --> 00:01:57,544
solutions for mobile devices. 
There are some demos online that I

26
00:01:57,544 --> 00:02:01,656
encourage you to actually watch. 
because this really doesn't show the

27
00:02:01,656 --> 00:02:04,620
idea. 
This is of course a static version.

28
00:02:04,620 --> 00:02:08,930
But for example one of the demos is you 
take a, an oil painting here.

29
00:02:08,930 --> 00:02:13,208
And you acquire it in such a way that if 
you show it on a tablet and you move it,

30
00:02:13,208 --> 00:02:19,140
it will actually exactly look like the 
original oil painting.

31
00:02:19,140 --> 00:02:22,450
So you get the illusion that you have 
actually the oil painting in the hand.

32
00:02:22,450 --> 00:02:26,550
So if the light changes, the vis, 
visualization will change.

33
00:02:26,550 --> 00:02:29,430
If you turn the tablet, the visualization 
will change.

34
00:02:29,430 --> 00:02:34,326
so then, what is quite stunning and I 
suggest you actually watch it, similarly,

35
00:02:34,326 --> 00:02:39,390
there is another demo which deals with 
stain glasses.

36
00:02:39,390 --> 00:02:42,909
So stain glasses are very interesting art 
objects, but very difficult to render on

37
00:02:42,909 --> 00:02:46,282
the internet. 
And so here we will stimulate the stained

38
00:02:46,282 --> 00:02:49,762
glass, so if you have a tablet in your 
hand and you move it, it looks like if

39
00:02:49,762 --> 00:02:56,768
you had, stained glass in your hand. 
So the tools we use is, we use

40
00:02:56,768 --> 00:03:01,750
traditional cameras, but we also use 
so-called light field cameras.

41
00:03:01,750 --> 00:03:03,930
You might have heard of the light 
[INAUDIBLE] for example.

42
00:03:03,930 --> 00:03:07,170
That's a new generation of camera. 
It's extremely interesting.

43
00:03:07,170 --> 00:03:11,885
And so we need to fully understand light 
transport theory and [INAUDIBLE].

44
00:03:11,885 --> 00:03:15,420
Which uses sparse recovery methods or 
compress sensing.

45
00:03:15,420 --> 00:03:20,798
The website of the project is given here. 
And as I indicated there is YouTube demo

46
00:03:20,798 --> 00:03:25,550
that shows quite realistically the demos 
that we're discussed just in a minute

47
00:03:25,550 --> 00:03:31,415
ago. 
So, next project is about wireless sensor

48
00:03:31,415 --> 00:03:36,510
networks, in particular about monitoring 
visually in a wireless sensor network.

49
00:03:36,510 --> 00:03:39,230
So, sensor networks have deployed for 
many years.

50
00:03:39,230 --> 00:03:43,860
We have large projects here, in the lab, 
on environmental monitoring.

51
00:03:43,860 --> 00:03:48,014
And the current generation is actually 
equipped with camera, and then you have a

52
00:03:48,014 --> 00:03:51,880
problem of compression, of 
representation.

53
00:03:51,880 --> 00:03:55,479
So even though the trend is towards 
smaller and smaller devices, they are

54
00:03:55,479 --> 00:03:59,255
still power hungry and in particular if 
you have a sophisticated camera, the

55
00:03:59,255 --> 00:04:03,326
number of images or the number of pixels 
that is generated might actually overwelm

56
00:04:03,326 --> 00:04:08,982
the power budget of the system. 
And so Dr.

57
00:04:08,982 --> 00:04:13,140
Zichong Chen, who finished his PhD here, 
and did his post doc, together with

58
00:04:13,140 --> 00:04:17,430
Guillermo Barrenetxea, are looking at 
creating large scale sensor networks

59
00:04:17,430 --> 00:04:22,289
equipped with cameras that are energy 
efficient.

60
00:04:24,050 --> 00:04:28,620
So why do we want images? 
Well, here are a few examples.

61
00:04:28,620 --> 00:04:33,250
This is from I think a Berkeley project 
about, monitoring birds' nest.

62
00:04:33,250 --> 00:04:36,890
Unfortunately a snake is showing up an 
he's actually eating all the eggs in the

63
00:04:36,890 --> 00:04:40,380
nest. 
So that's, monitoring for why life

64
00:04:40,380 --> 00:04:43,455
protection. 
Here is an example from the Swiss Alps

65
00:04:43,455 --> 00:04:48,117
monitoring for for avalanche detection. 
Here is also an example from the Swiss

66
00:04:48,117 --> 00:04:51,850
Alps, its monitoring to see weather 
conditions.

67
00:04:51,850 --> 00:04:55,810
And, finally, here is monitoring networks 
that is installed on the PFL campus.

68
00:04:55,810 --> 00:05:00,990
In all these cases you have many cameras 
using small communication devices.

69
00:05:00,990 --> 00:05:04,690
And so compression and representation is 
extremely critical.

70
00:05:06,630 --> 00:05:12,619
So there are a number of results that you 
can find in the thesis of Dr.

71
00:05:12,619 --> 00:05:17,437
Chi Chong Chang, given here in this, 
website, and essentially the idea is

72
00:05:17,437 --> 00:05:21,963
that, cameras can help each other to 
reduce the amount of information that

73
00:05:21,963 --> 00:05:26,197
actually has to be sent to the base 
station or into the cloud, for doing

74
00:05:26,197 --> 00:05:32,868
efficient monitoring. 
So a long with signal processing today,

75
00:05:32,868 --> 00:05:36,275
actually it's moving to single processing 
on graphs.

76
00:05:36,275 --> 00:05:39,815
I don't have to explain to you the 
importance, for example, of social

77
00:05:39,815 --> 00:05:43,514
networks. 
And so Pedro Pinto who was involved here

78
00:05:43,514 --> 00:05:47,610
in the class and is a post doc in the 
lab, together with Patrick Thiran, has

79
00:05:47,610 --> 00:05:52,230
worked on the problem of source 
localization.

80
00:05:52,230 --> 00:05:57,158
So you have some graph here, let's say 
social network and somebody launches a

81
00:05:57,158 --> 00:06:00,850
rumor. 
Here is the source and the rumor gets

82
00:06:00,850 --> 00:06:05,252
forwarded along the edges of the graph at 
different times and you have some

83
00:06:05,252 --> 00:06:12,570
observers, say green nodes here, that 
receives a rumor at some instant of time.

84
00:06:12,570 --> 00:06:15,770
They know where the rumor comes from, you 
know who told you the gossip, and you

85
00:06:15,770 --> 00:06:19,976
know when you got the gossip information. 
So, the question is, you know the

86
00:06:19,976 --> 00:06:23,000
structure of the graph, or you have an 
approximation of the structure of the

87
00:06:23,000 --> 00:06:26,380
graphs. 
You have these observations.

88
00:06:26,380 --> 00:06:30,290
Can you figure out who actually, spreads 
the rumor first.

89
00:06:30,290 --> 00:06:33,450
It turns out this has, an interesting 
solution.

90
00:06:33,450 --> 00:06:38,349
And using only few observers, about 20%, 
you can achieve a very high accuracy in

91
00:06:38,349 --> 00:06:43,510
finding the source of a rumor on a large 
scale network.

92
00:06:43,510 --> 00:06:46,930
And there are many interesting questions 
here, to pursue in this source

93
00:06:46,930 --> 00:06:51,291
localization in social networks. 
And there was a paper that came out last

94
00:06:51,291 --> 00:06:55,452
year, Locating the Source of Diffusion in 
Large-Scale Networks that had quite a bit

95
00:06:55,452 --> 00:06:59,101
of impact. 
The project is actually funded by the

96
00:06:59,101 --> 00:07:02,600
Bill and Melinda Gates Foundation. 
The reason is that one of the

97
00:07:02,600 --> 00:07:05,650
applications is to monitor health 
problems.

98
00:07:05,650 --> 00:07:10,642
For example, here is a map of Cholera 
outbreak in Africa, and the map shows the

99
00:07:10,642 --> 00:07:15,370
river network. 
Cholera is a water born disease, and so,

100
00:07:15,370 --> 00:07:19,735
typically Cholera will actually diffuse 
along waterways.

101
00:07:19,735 --> 00:07:23,767
But you know when people fell sick at 
certain locations and then you can infer

102
00:07:23,767 --> 00:07:27,410
the source of the actual Cholera 
outbreak.

103
00:07:27,410 --> 00:07:31,370
There is another example here, which is a 
simulation of, if you had to figure out

104
00:07:31,370 --> 00:07:35,150
if there was some pollution or attack on 
the New York subway, and if you could

105
00:07:35,150 --> 00:07:39,230
figure out knowing the network of the New 
York subway and when you start detecting

106
00:07:39,230 --> 00:07:45,750
the problems where the source of the 
problem actually was.

107
00:07:47,310 --> 00:07:51,279
The next project is on sampling, so we 
have worked on a new theory of sampling

108
00:07:51,279 --> 00:07:54,681
here called Finite Rate of Innovation 
Sampling, and it is used in

109
00:07:54,681 --> 00:07:58,650
communications problems, and in 
monitoring problems to reduce the number

110
00:07:58,650 --> 00:08:03,957
of samples being transmitted or acquired. 
Dr.

111
00:08:03,957 --> 00:08:08,446
Freris, who is a senior scientist with 
doctoral students and MS assistants, are

112
00:08:08,446 --> 00:08:12,868
actually working on doing ECG monitoring 
at very low power for wireless health

113
00:08:12,868 --> 00:08:17,002
monitoring. 
So here is a block diagram, it's

114
00:08:17,002 --> 00:08:21,032
relatively complicated so let me not get 
into this, but it uses some fairly

115
00:08:21,032 --> 00:08:25,186
sophisticated techniques to reduce the 
sampling rate so as to reduce the energy

116
00:08:25,186 --> 00:08:33,080
consumption on these wireless devices. 
So, this project is actually sponsored by

117
00:08:33,080 --> 00:08:40,165
somebody well known, Qualcomm, interested 
in the theory of sampling.

118
00:08:40,165 --> 00:08:43,637
And the extension here for this 
particular project has been

119
00:08:43,637 --> 00:08:49,810
generalization of the initial finite rate 
of innovation sampling methodology.

120
00:08:49,810 --> 00:08:53,730
To get better compression, and better 
modelization of the signals.

121
00:08:53,730 --> 00:08:57,825
So here we have the ECG signal, and then 
there is sophisticated models that

122
00:08:57,825 --> 00:09:02,795
allows, to take very few parameters, to 
model the ECG signal.

123
00:09:02,795 --> 00:09:06,500
There are a number of papers here, the 
initial paper on finite rate of

124
00:09:06,500 --> 00:09:10,920
innovation sampling is this 2002 paper, 
and the number of recent papers have done

125
00:09:10,920 --> 00:09:16,940
extension to this theory. 
So if you like sampling I welcome you to

126
00:09:16,940 --> 00:09:22,750
actually read up on this stuff, it's one 
of my favorite research topics.

127
00:09:24,880 --> 00:09:28,845
When we talk about sampling already in 
sensor networks we have mentioned that

128
00:09:28,845 --> 00:09:34,356
placing a sensor is like taking a sample. 
And so that spatial sampling, now if you

129
00:09:34,356 --> 00:09:38,596
do spatial sampling, you can also use 
mobile sensors and Dr.

130
00:09:38,596 --> 00:09:43,152
Unnikrishnan here, a post doc in the lab, 
has worked on this or generalization of

131
00:09:43,152 --> 00:09:47,708
the theory of sampling when you have 
mobile sensors that can actually go over

132
00:09:47,708 --> 00:09:56,360
a field in an arbitrary fashion. 
Then you maybe show this in an example.

133
00:09:56,360 --> 00:10:00,203
It's again a temperature monitoring 
example here on the EPFL campus, or you

134
00:10:00,203 --> 00:10:03,971
have buildings. 
You have that open space between

135
00:10:03,971 --> 00:10:07,176
buildings. 
Those buildings are, of course, hot.

136
00:10:07,176 --> 00:10:11,327
The open space are cool. 
And you would like to have monitoring of

137
00:10:11,327 --> 00:10:16,699
this temperature field not with static 
spatial sensors, but with people running

138
00:10:16,699 --> 00:10:23,960
around, having a thermal meter let's say 
on their mobile phone.

139
00:10:23,960 --> 00:10:27,854
And the question is, how accurate can you 
actually measure temperature using a

140
00:10:27,854 --> 00:10:32,240
device like this? 
And so, this is being done actually for

141
00:10:32,240 --> 00:10:37,612
pollution monitoring in the city of 
Lausanne so there's some equipment put on

142
00:10:37,612 --> 00:10:44,827
buses to measure pollution parameters. 
And what we do here is we try to develop

143
00:10:44,827 --> 00:10:49,909
a theory of how good you can sample when 
you have these mobile sensors going over

144
00:10:49,909 --> 00:10:57,310
a surface and measuring a field. 
The results are very mathematical but are

145
00:10:57,310 --> 00:11:00,660
interesting because our non-trivial 
extension of sampling theory through

146
00:11:00,660 --> 00:11:04,614
multiple dimensions. 
And a few papers are mentioned here if

147
00:11:04,614 --> 00:11:11,010
you are interested in more detail. 
The next project is about a new way of

148
00:11:11,010 --> 00:11:15,470
doing image acquisition. 
So in this class, we have seen sampling

149
00:11:15,470 --> 00:11:20,484
and we have seen quantization. 
And when we do quantization typically we

150
00:11:20,484 --> 00:11:25,460
say, let's take [UNKNOWN] samples and 
then take as many bits as possible.

151
00:11:25,460 --> 00:11:30,920
Let's say eight bits for speech, 12 bits 
for images 24 bits maybe for audio,

152
00:11:30,920 --> 00:11:35,310
etcetera. 
Now here we took the extreme other

153
00:11:35,310 --> 00:11:40,332
example we said lets build an image 
sensor that has many, many, many pixels

154
00:11:40,332 --> 00:11:46,690
but the pixels only detect either a 
enough light or not.

155
00:11:46,690 --> 00:11:49,080
So the pixels are actually binary 
detectors.

156
00:11:49,080 --> 00:11:54,610
And so you have a light intensity here. 
Which changes over space.

157
00:11:54,610 --> 00:11:58,200
You have a lens that smooths the light 
intensity.

158
00:11:58,200 --> 00:12:01,110
So what reaches the camera is this smooth 
curve here.

159
00:12:01,110 --> 00:12:05,060
And this smooth curve you sample very, 
very, very finely.

160
00:12:05,060 --> 00:12:09,410
But you only decide if it's above or 
below a certain threshold.

161
00:12:09,410 --> 00:12:13,480
So the sensor only generates a sequence 
of binary digits.

162
00:12:13,480 --> 00:12:17,850
So that's the imaging model. 
And this has been studied by Dr.

163
00:12:17,850 --> 00:12:22,008
Feng Yang, did his PhD thesis on this, is 
now a post-doc working on this project,

164
00:12:22,008 --> 00:12:27,370
and a whole slew of other people. 
This was a very extensive project.

165
00:12:27,370 --> 00:12:32,765
And what is interesting is that this new 
way of acquiring images, for example,

166
00:12:32,765 --> 00:12:40,089
allows to do high dynamic range imaging. 
Here is a simulation of a high dynamic

167
00:12:40,089 --> 00:12:45,870
range image in a much easier way than 
with conventional cameras.

168
00:12:45,870 --> 00:12:49,736
That's one advantage. 
Another one is that you can have very,

169
00:12:49,736 --> 00:12:52,615
very cheap sensors. 
So here's an example of one that was

170
00:12:52,615 --> 00:12:56,800
built in the lab. 
And then, you take many, many frames.

171
00:12:56,800 --> 00:12:59,832
They are extremely noisy. 
If they look noisy, they are simply

172
00:12:59,832 --> 00:13:03,660
binary, so you only have zeroes and ones, 
but you have enough of these, and you do

173
00:13:03,660 --> 00:13:09,955
an optimal reconstruction method. 
You actually can recognize here, the logo

174
00:13:09,955 --> 00:13:15,026
of EPFL. 
There are publications here that you are

175
00:13:15,026 --> 00:13:19,380
welcome to look up. 
And the thesis is online.

176
00:13:19,380 --> 00:13:23,230
Last but not least Rambus silicon valley 
company, actually works with us on this

177
00:13:23,230 --> 00:13:28,620
and has acquired some of the technologies 
that was developed in this project.

178
00:13:30,790 --> 00:13:33,600
And old classic is trying to predict the 
stock market.

179
00:13:33,600 --> 00:13:38,438
So, we gave it another shot. 
so Lionel Coulot did his PhD thesis, was

180
00:13:38,438 --> 00:13:43,100
co-advised with Peter Bossaerts who is at 
Caltech.

181
00:13:43,100 --> 00:13:47,316
And we were trying to understand if 
methods from information theory would

182
00:13:47,316 --> 00:13:51,668
allow to predict models for the stock 
market, and that requires statistical

183
00:13:51,668 --> 00:13:56,234
models for what the stock market might 
be.

184
00:13:56,234 --> 00:14:00,264
And what is interesting is that you have 
to decide between very sophisticated

185
00:14:00,264 --> 00:14:04,294
models that might be overkill and are 
hard to estimate, and very simple models

186
00:14:04,294 --> 00:14:08,200
which might be too simplistic, but which 
might be very robust to things that

187
00:14:08,200 --> 00:14:15,118
happen in the stock market. 
And, in the end we used coding theory and

188
00:14:15,118 --> 00:14:20,082
classic algorithmic methods like dynamic 
programming to come up with a method that

189
00:14:20,082 --> 00:14:24,434
decides what is the correct model at 
every time of, the observation of the

190
00:14:24,434 --> 00:14:31,690
stock market. 
So I'm just going to show a picture.

191
00:14:31,690 --> 00:14:35,939
And the picture is, is a value on the 
stock market.

192
00:14:35,939 --> 00:14:40,412
And the question is, can you detect if 
the stock market is in a bear market or a

193
00:14:40,412 --> 00:14:44,562
bull market? 
So when the stock market goes up it's,

194
00:14:44,562 --> 00:14:48,130
called bull market. 
If it goes down, it's a bear market.

195
00:14:48,130 --> 00:14:53,180
What is very hard is to decide by 
watching every day what's happening.

196
00:14:53,180 --> 00:14:56,582
If currently the trend is going up or the 
trend is going down and you need to do

197
00:14:56,582 --> 00:15:01,857
this with an online algorithm. 
Okay, you cannot look into the future and

198
00:15:01,857 --> 00:15:06,543
this method developed by Lionel allows to 
do a model fitting and to very quickly

199
00:15:06,543 --> 00:15:14,050
detect when the stock market changes from 
a bull market to a bar, bear market.

200
00:15:15,710 --> 00:15:21,970
The thesis online and this was sponsored 
by, as you may guess, by a bank.

201
00:15:21,970 --> 00:15:26,041
And the results are interesting, but we 
are still having a regular day job so you

202
00:15:26,041 --> 00:15:29,404
can guess that the method is not 
completely fool proof to predict the

203
00:15:29,404 --> 00:15:33,590
stock market. 
But the methods, the algorithms, and the

204
00:15:33,590 --> 00:15:39,296
theory behind it is quite cool. 
The next few projects are so called

205
00:15:39,296 --> 00:15:42,708
inverse problems. 
So inverse problems are problems where

206
00:15:42,708 --> 00:15:46,236
you have some measurements but the 
measurements do not describe the signal

207
00:15:46,236 --> 00:15:50,649
you're interested in. 
But some indirect measurement of the

208
00:15:50,649 --> 00:15:56,210
signal, so you try to invert the system 
to go back to the original signal.

209
00:15:56,210 --> 00:15:59,801
You all know about computerized 
tomography, a medical image method, where

210
00:15:59,801 --> 00:16:03,660
you can see inside the body without 
really going there.

211
00:16:03,660 --> 00:16:08,004
And that's a typical inverse problem. 
Here we are interested in inverse

212
00:16:08,004 --> 00:16:13,110
problems in environmental monitoring. 
So, the first example is diffusion

213
00:16:13,110 --> 00:16:16,545
equation. 
And we have a physical phenomena, for

214
00:16:16,545 --> 00:16:22,150
example temperature has been discussed, 
or atmospheric dispersal of pollution.

215
00:16:22,150 --> 00:16:26,840
We want to measure the field at locations 
where we can put sensors, and the goal is

216
00:16:26,840 --> 00:16:32,110
to find where are the sources, for 
example, of pollution.

217
00:16:32,110 --> 00:16:36,867
Now this is a hard problem because, you 
have to model how, for example, pollution

218
00:16:36,867 --> 00:16:40,867
is being diffused. 
That depends on weather patterns and so

219
00:16:40,867 --> 00:16:44,037
on. 
But the tools we are using are typical

220
00:16:44,037 --> 00:16:48,945
signal processing tools, for analysis. 
Sampling theory for exemplifying finite

221
00:16:48,945 --> 00:16:52,245
rate of innovation sampling or 
compressive sensing, that has also been

222
00:16:52,245 --> 00:16:56,860
mentioned earlier. 
Let's look at the picture.

223
00:16:56,860 --> 00:17:01,488
That's a very simple example of this. 
Assume you have two smokestacks and

224
00:17:01,488 --> 00:17:06,992
inside a factory compound, and the 
smokestacks produce pollution which

225
00:17:06,992 --> 00:17:12,210
changes every day. 
You don't know how much pollution is

226
00:17:12,210 --> 00:17:16,820
being released, and you're working for an 
environmental monitoring agency.

227
00:17:16,820 --> 00:17:21,644
You put sensors outside of the compound 
and you measure what arrives, in terms of

228
00:17:21,644 --> 00:17:26,569
pollution, at these sensors. 
And the goal is to figure out if what

229
00:17:26,569 --> 00:17:29,985
came out of smoke stack was within the 
bounds allowed, lets say by z,

230
00:17:29,985 --> 00:17:35,966
Environmental Protection Agency. 
So this is a interesting and non-trivial

231
00:17:35,966 --> 00:17:40,776
problem but there are some interesting 
results that were produced by Yuri

232
00:17:40,776 --> 00:17:46,104
Ranieri, whom you all know because he was 
the famous Master Chief assistant for the

233
00:17:46,104 --> 00:17:53,992
BSB class. 
So we are able to recover sparse sources

234
00:17:53,992 --> 00:17:58,378
using this inversion method. 
we use this finite rate of innovation

235
00:17:58,378 --> 00:18:02,630
sampling techniques to actually do it. 
And here we is a list of publications

236
00:18:02,630 --> 00:18:10,180
that came out of this research. 
This problem is also an inverse problem.

237
00:18:10,180 --> 00:18:13,740
It's a Fukushima inverse problem. 
It is a PhD project of Marta

238
00:18:13,740 --> 00:18:19,152
Martinez-Camara, and a few other of us 
are involved in this, and we collaborate

239
00:18:19,152 --> 00:18:26,180
with a specialist Andreas Stohl. 
Who is a specialist of monitoring of

240
00:18:26,180 --> 00:18:30,610
radioactive diffusion. 
So what we like to do is figure out how

241
00:18:30,610 --> 00:18:36,280
much radionuclides were actually released 
in Fukushima at the time of the of the

242
00:18:36,280 --> 00:18:42,560
nuclear accident at Fukushima. 
We have only very few sensors, they are

243
00:18:42,560 --> 00:18:46,060
located around the world very far away 
from Fukushima.

244
00:18:46,060 --> 00:18:51,085
And the question is, is it possible from 
these few measurements around the world

245
00:18:51,085 --> 00:18:55,885
taken later, to invert the entire process 
as I diffuse the initial release of

246
00:18:55,885 --> 00:19:03,150
radioactive material into the atmosphere. 
What tools are we using?

247
00:19:03,150 --> 00:19:06,300
Sparse regularizations, so that's 
compressed sensing.

248
00:19:06,300 --> 00:19:09,663
And we need to using atmospheric 
dispersion model to understand how

249
00:19:09,663 --> 00:19:13,616
radioactive material from from Fukushima 
was actually transported across the

250
00:19:13,616 --> 00:19:19,568
world. 
So one result that we have and which is

251
00:19:19,568 --> 00:19:24,260
very interesting is we were able to 
estimate the emission of Xenons, that's

252
00:19:24,260 --> 00:19:28,884
radioactive gas that was released at the 
time of explosions at Fukushima, went up

253
00:19:28,884 --> 00:19:36,798
into the atmosphere, was transported by 
weather patterns all over the world.

254
00:19:36,798 --> 00:19:41,218
And from the measurements all over the 
world, we were able to pinpoint exactly

255
00:19:41,218 --> 00:19:48,560
when the Xenon was released, and how much 
Xenon was released into the atmosphere.

256
00:19:48,560 --> 00:19:52,272
And it turns out we actually know the 
total amount of Xenon that was released,

257
00:19:52,272 --> 00:19:57,960
because after the accident no Xenon was 
actually left in the nuclear power plant.

258
00:19:59,890 --> 00:20:03,754
Currently we're trying to go beyond this 
and estimate the Cesium release, but that

259
00:20:03,754 --> 00:20:08,518
turns out to be a harder problem. 
The paper that describes this will be

260
00:20:08,518 --> 00:20:13,860
published ICASSP this year and is 
available online here in infoscience.

261
00:20:16,390 --> 00:20:20,614
Last but not least is a project we call, 
"Can One Hear the Shape of a Room?" It's

262
00:20:20,614 --> 00:20:24,574
a PhD project of Ivan Dokmanic and 
several other people in the lab, in

263
00:20:24,574 --> 00:20:31,110
particular, Reza Parhizkar, Andreaz 
Walther, have worked on this.

264
00:20:31,110 --> 00:20:33,806
And also we have a collaboration with Yue 
Lu.

265
00:20:33,806 --> 00:20:38,650
He's now with Harvard. 
Now you know about this problem because,

266
00:20:38,650 --> 00:20:43,980
Ivan gave module 512 about gear 
dereverberation, echo cancellation.

267
00:20:43,980 --> 00:20:48,628
And, uh,the next step is to say, if I 
listen to echoes, can I actually

268
00:20:48,628 --> 00:20:55,008
understand what is a room shape? 
So if I know the room shape, then I know

269
00:20:55,008 --> 00:20:59,898
how to generate the echoes. 
But if you give me the echoes, can I know

270
00:20:59,898 --> 00:21:03,165
the room shape? 
It's a classic inverse problem, very cute

271
00:21:03,165 --> 00:21:05,981
one. 
And we usually explain it by saying,

272
00:21:05,981 --> 00:21:10,470
let's say you enter a room, you're 
blindfolded.

273
00:21:10,470 --> 00:21:13,820
And so you don't see the room at all. 
You snap your finger.

274
00:21:13,820 --> 00:21:19,280
You therefore elicit echoes, you listen 
very carefully to the echoes.

275
00:21:19,280 --> 00:21:22,460
Can you exactly see or hear the shape of 
the room?

276
00:21:24,390 --> 00:21:27,756
Now this has a beautiful theory, which we 
won't have time to really explain, but

277
00:21:27,756 --> 00:21:31,660
that you can read up about because it's 
published material.

278
00:21:31,660 --> 00:21:36,490
But if you have a source or receiver you 
have a direct pass between the source and

279
00:21:36,490 --> 00:21:41,540
the receiver, and you have echoes given 
by the walls.

280
00:21:41,540 --> 00:21:45,895
The echoes given by the walls correspond 
to so called mirror or image sources, so

281
00:21:45,895 --> 00:21:49,665
this is the same as if you had a source 
here and the sound would have gone

282
00:21:49,665 --> 00:21:54,666
straight here. 
So if you can locate all these image

283
00:21:54,666 --> 00:21:58,750
sources, then you can actually locate the 
room.

284
00:21:58,750 --> 00:22:02,772
The walls, therefore the room. 
And this is, you know, in principal

285
00:22:02,772 --> 00:22:08,590
do-able the question was is it always 
true that this can be done?

286
00:22:08,590 --> 00:22:11,501
And is it also realistic to do it in 
practice?

287
00:22:11,501 --> 00:22:15,659
So, here are examples of a system with 
five microforms, you have one source five

288
00:22:15,659 --> 00:22:19,466
microforms. 
You have somebody snap his finger and you

289
00:22:19,466 --> 00:22:23,498
have the echos related to the walls and 
you see there is a complexity which is,

290
00:22:23,498 --> 00:22:27,467
these echos come in random orders because 
different walls are at different

291
00:22:27,467 --> 00:22:33,798
distances of the microphone. 
And the question is, can we find out the

292
00:22:33,798 --> 00:22:37,780
shape from a set of measurements as we 
see here?

293
00:22:37,780 --> 00:22:44,050
How many measurements do we need? 
Can we have a robust algorithm?

294
00:22:44,050 --> 00:22:48,490
So the answer is summarized in, yes we 
can.

295
00:22:48,490 --> 00:22:52,157
And there are some experiments we did, 
both at the labs.

296
00:22:52,157 --> 00:22:56,122
So this is one of our seminar rooms we 
created a, an artificial wall here to

297
00:22:56,122 --> 00:23:01,330
have different shapes of rooms. 
So this is a typical shape of room.

298
00:23:01,330 --> 00:23:06,082
Then in this case, with five microphone 
and one source, we were able to estimate

299
00:23:06,082 --> 00:23:12,290
the size, shape of the room very 
accurately to more, better than 1%.

300
00:23:12,290 --> 00:23:16,996
And once we had this, we said, well, 
let's see how robust this is.

301
00:23:16,996 --> 00:23:21,284
We went to Lausanne Cathedral and that's 
actually a foyer of the Lausanne

302
00:23:21,284 --> 00:23:25,840
Cathedral, which is not at all needing 
the assumptions of the algorithms that

303
00:23:25,840 --> 00:23:31,677
I've described very briefly here. 
And it was still possible to see the

304
00:23:31,677 --> 00:23:36,420
major refractors, meaning the major walls 
here in the Lausanne Cathedral.

305
00:23:36,420 --> 00:23:40,200
And so the answer is yes, one can hear 
the shape of a room.

306
00:23:40,200 --> 00:23:44,840
And you can visit Ivan's web page to see 
more details.

307
00:23:46,980 --> 00:23:50,518
Now these were just a selection of 
projects, of works that is being done by

308
00:23:50,518 --> 00:23:55,100
PhDs and post-docs and senior researchers 
in the lab.

309
00:23:55,100 --> 00:23:59,126
Please go to the website, as that gives 
the entire portfolio of research here of

310
00:23:59,126 --> 00:24:02,010
what the lab is currently doing.