0:00:00.660,0:00:03.960
Module 10.2. [br]Some research projects that use the

0:00:03.960,0:00:07.770
techniques you have learned in the [br]digital signal processing class.

0:00:09.050,0:00:11.890
We're going to talk about some current [br]research in the lab.

0:00:11.890,0:00:15.982
There is a whole slew of them, and it's a [br]selection here of interesting research

0:00:15.982,0:00:21.642
projects that we can briefly discuss. [br]The first one is, eFacsimile is a project

0:00:21.642,0:00:26.670
on art work acquisition. [br]The second one is about signal processing

0:00:26.670,0:00:30.170
in sensor networks. [br]Then there is a network science result on

0:00:30.170,0:00:35.020
source localization put in graphs. [br]Then we talk about sampling result, so

0:00:35.020,0:00:38.690
called finite rate of innovation [br]sampling.

0:00:38.690,0:00:42.472
Then we talk again about sampling, that [br]of physical fields, using some new

0:00:42.472,0:00:47.260
techniques for sampling. [br]Then, we have a project on image

0:00:47.260,0:00:52.720
acquisition where we change the sensors [br]used in acquiring images.

0:00:52.720,0:00:55.270
Then, an old classic, predicting the [br]stock market.

0:00:56.550,0:01:00.587
Then, we talk about inverse problems. [br]The three next projects are actually

0:01:00.587,0:01:04.209
inverse problems. [br]The first one is on the diffusion

0:01:04.209,0:01:08.363
equation, the second one is trying to [br]understand the nuclear fall out from

0:01:08.363,0:01:14.400
Fukushima, and last but not least is an [br]inverse problem in acoustics.

0:01:17.460,0:01:20.480
The eFacsimile project. [br]This is a project that we do together

0:01:20.480,0:01:26.800
with Google to try to improve how artwork [br]is represented on the internet.

0:01:26.800,0:01:31.910
it's lead by [INAUDIBLE] researcher Loic [br]Baboulaz and several PhD students are

0:01:31.910,0:01:36.117
involved in this. [br]The questions are how to capture,

0:01:36.117,0:01:40.670
represent, and render artwork as well as [br]possible.

0:01:40.670,0:01:45.415
And to do this, we need some advanced [br]techniques on relighting, manipulation

0:01:45.415,0:01:50.452
of, the so called, light fields that is [br]acquired, and potentially high resolution

0:01:50.452,0:01:57.544
solutions for mobile devices. [br]There are some demos online that I

0:01:57.544,0:02:01.656
encourage you to actually watch. [br]because this really doesn't show the

0:02:01.656,0:02:04.620
idea. [br]This is of course a static version.

0:02:04.620,0:02:08.930
But for example one of the demos is you [br]take a, an oil painting here.

0:02:08.930,0:02:13.208
And you acquire it in such a way that if [br]you show it on a tablet and you move it,

0:02:13.208,0:02:19.140
it will actually exactly look like the [br]original oil painting.

0:02:19.140,0:02:22.450
So you get the illusion that you have [br]actually the oil painting in the hand.

0:02:22.450,0:02:26.550
So if the light changes, the vis, [br]visualization will change.

0:02:26.550,0:02:29.430
If you turn the tablet, the visualization [br]will change.

0:02:29.430,0:02:34.326
so then, what is quite stunning and I [br]suggest you actually watch it, similarly,

0:02:34.326,0:02:39.390
there is another demo which deals with [br]stain glasses.

0:02:39.390,0:02:42.909
So stain glasses are very interesting art [br]objects, but very difficult to render on

0:02:42.909,0:02:46.282
the internet. [br]And so here we will stimulate the stained

0:02:46.282,0:02:49.762
glass, so if you have a tablet in your [br]hand and you move it, it looks like if

0:02:49.762,0:02:56.768
you had, stained glass in your hand. [br]So the tools we use is, we use

0:02:56.768,0:03:01.750
traditional cameras, but we also use [br]so-called light field cameras.

0:03:01.750,0:03:03.930
You might have heard of the light [br][INAUDIBLE] for example.

0:03:03.930,0:03:07.170
That's a new generation of camera. [br]It's extremely interesting.

0:03:07.170,0:03:11.885
And so we need to fully understand light [br]transport theory and [INAUDIBLE].

0:03:11.885,0:03:15.420
Which uses sparse recovery methods or [br]compress sensing.

0:03:15.420,0:03:20.798
The website of the project is given here. [br]And as I indicated there is YouTube demo

0:03:20.798,0:03:25.550
that shows quite realistically the demos [br]that we're discussed just in a minute

0:03:25.550,0:03:31.415
ago. [br]So, next project is about wireless sensor

0:03:31.415,0:03:36.510
networks, in particular about monitoring [br]visually in a wireless sensor network.

0:03:36.510,0:03:39.230
So, sensor networks have deployed for [br]many years.

0:03:39.230,0:03:43.860
We have large projects here, in the lab, [br]on environmental monitoring.

0:03:43.860,0:03:48.014
And the current generation is actually [br]equipped with camera, and then you have a

0:03:48.014,0:03:51.880
problem of compression, of [br]representation.

0:03:51.880,0:03:55.479
So even though the trend is towards [br]smaller and smaller devices, they are

0:03:55.479,0:03:59.255
still power hungry and in particular if [br]you have a sophisticated camera, the

0:03:59.255,0:04:03.326
number of images or the number of pixels [br]that is generated might actually overwelm

0:04:03.326,0:04:08.982
the power budget of the system. [br]And so Dr.

0:04:08.982,0:04:13.140
Zichong Chen, who finished his PhD here, [br]and did his post doc, together with

0:04:13.140,0:04:17.430
Guillermo Barrenetxea, are looking at [br]creating large scale sensor networks

0:04:17.430,0:04:22.289
equipped with cameras that are energy [br]efficient.

0:04:24.050,0:04:28.620
So why do we want images? [br]Well, here are a few examples.

0:04:28.620,0:04:33.250
This is from I think a Berkeley project [br]about, monitoring birds' nest.

0:04:33.250,0:04:36.890
Unfortunately a snake is showing up an [br]he's actually eating all the eggs in the

0:04:36.890,0:04:40.380
nest. [br]So that's, monitoring for why life

0:04:40.380,0:04:43.455
protection. [br]Here is an example from the Swiss Alps

0:04:43.455,0:04:48.117
monitoring for for avalanche detection. [br]Here is also an example from the Swiss

0:04:48.117,0:04:51.850
Alps, its monitoring to see weather [br]conditions.

0:04:51.850,0:04:55.810
And, finally, here is monitoring networks [br]that is installed on the PFL campus.

0:04:55.810,0:05:00.990
In all these cases you have many cameras [br]using small communication devices.

0:05:00.990,0:05:04.690
And so compression and representation is [br]extremely critical.

0:05:06.630,0:05:12.619
So there are a number of results that you [br]can find in the thesis of Dr.

0:05:12.619,0:05:17.437
Chi Chong Chang, given here in this, [br]website, and essentially the idea is

0:05:17.437,0:05:21.963
that, cameras can help each other to [br]reduce the amount of information that

0:05:21.963,0:05:26.197
actually has to be sent to the base [br]station or into the cloud, for doing

0:05:26.197,0:05:32.868
efficient monitoring. [br]So a long with signal processing today,

0:05:32.868,0:05:36.275
actually it's moving to single processing [br]on graphs.

0:05:36.275,0:05:39.815
I don't have to explain to you the [br]importance, for example, of social

0:05:39.815,0:05:43.514
networks. [br]And so Pedro Pinto who was involved here

0:05:43.514,0:05:47.610
in the class and is a post doc in the [br]lab, together with Patrick Thiran, has

0:05:47.610,0:05:52.230
worked on the problem of source [br]localization.

0:05:52.230,0:05:57.158
So you have some graph here, let's say [br]social network and somebody launches a

0:05:57.158,0:06:00.850
rumor. [br]Here is the source and the rumor gets

0:06:00.850,0:06:05.252
forwarded along the edges of the graph at [br]different times and you have some

0:06:05.252,0:06:12.570
observers, say green nodes here, that [br]receives a rumor at some instant of time.

0:06:12.570,0:06:15.770
They know where the rumor comes from, you [br]know who told you the gossip, and you

0:06:15.770,0:06:19.976
know when you got the gossip information. [br]So, the question is, you know the

0:06:19.976,0:06:23.000
structure of the graph, or you have an [br]approximation of the structure of the

0:06:23.000,0:06:26.380
graphs. [br]You have these observations.

0:06:26.380,0:06:30.290
Can you figure out who actually, spreads [br]the rumor first.

0:06:30.290,0:06:33.450
It turns out this has, an interesting [br]solution.

0:06:33.450,0:06:38.349
And using only few observers, about 20%, [br]you can achieve a very high accuracy in

0:06:38.349,0:06:43.510
finding the source of a rumor on a large [br]scale network.

0:06:43.510,0:06:46.930
And there are many interesting questions [br]here, to pursue in this source

0:06:46.930,0:06:51.291
localization in social networks. [br]And there was a paper that came out last

0:06:51.291,0:06:55.452
year, Locating the Source of Diffusion in [br]Large-Scale Networks that had quite a bit

0:06:55.452,0:06:59.101
of impact. [br]The project is actually funded by the

0:06:59.101,0:07:02.600
Bill and Melinda Gates Foundation. [br]The reason is that one of the

0:07:02.600,0:07:05.650
applications is to monitor health [br]problems.

0:07:05.650,0:07:10.642
For example, here is a map of Cholera [br]outbreak in Africa, and the map shows the

0:07:10.642,0:07:15.370
river network. [br]Cholera is a water born disease, and so,

0:07:15.370,0:07:19.735
typically Cholera will actually diffuse [br]along waterways.

0:07:19.735,0:07:23.767
But you know when people fell sick at [br]certain locations and then you can infer

0:07:23.767,0:07:27.410
the source of the actual Cholera [br]outbreak.

0:07:27.410,0:07:31.370
There is another example here, which is a [br]simulation of, if you had to figure out

0:07:31.370,0:07:35.150
if there was some pollution or attack on [br]the New York subway, and if you could

0:07:35.150,0:07:39.230
figure out knowing the network of the New [br]York subway and when you start detecting

0:07:39.230,0:07:45.750
the problems where the source of the [br]problem actually was.

0:07:47.310,0:07:51.279
The next project is on sampling, so we [br]have worked on a new theory of sampling

0:07:51.279,0:07:54.681
here called Finite Rate of Innovation [br]Sampling, and it is used in

0:07:54.681,0:07:58.650
communications problems, and in [br]monitoring problems to reduce the number

0:07:58.650,0:08:03.957
of samples being transmitted or acquired. [br]Dr.

0:08:03.957,0:08:08.446
Freris, who is a senior scientist with [br]doctoral students and MS assistants, are

0:08:08.446,0:08:12.868
actually working on doing ECG monitoring [br]at very low power for wireless health

0:08:12.868,0:08:17.002
monitoring. [br]So here is a block diagram, it's

0:08:17.002,0:08:21.032
relatively complicated so let me not get [br]into this, but it uses some fairly

0:08:21.032,0:08:25.186
sophisticated techniques to reduce the [br]sampling rate so as to reduce the energy

0:08:25.186,0:08:33.080
consumption on these wireless devices. [br]So, this project is actually sponsored by

0:08:33.080,0:08:40.165
somebody well known, Qualcomm, interested [br]in the theory of sampling.

0:08:40.165,0:08:43.637
And the extension here for this [br]particular project has been

0:08:43.637,0:08:49.810
generalization of the initial finite rate [br]of innovation sampling methodology.

0:08:49.810,0:08:53.730
To get better compression, and better [br]modelization of the signals.

0:08:53.730,0:08:57.825
So here we have the ECG signal, and then [br]there is sophisticated models that

0:08:57.825,0:09:02.795
allows, to take very few parameters, to [br]model the ECG signal.

0:09:02.795,0:09:06.500
There are a number of papers here, the [br]initial paper on finite rate of

0:09:06.500,0:09:10.920
innovation sampling is this 2002 paper, [br]and the number of recent papers have done

0:09:10.920,0:09:16.940
extension to this theory. [br]So if you like sampling I welcome you to

0:09:16.940,0:09:22.750
actually read up on this stuff, it's one [br]of my favorite research topics.

0:09:24.880,0:09:28.845
When we talk about sampling already in [br]sensor networks we have mentioned that

0:09:28.845,0:09:34.356
placing a sensor is like taking a sample. [br]And so that spatial sampling, now if you

0:09:34.356,0:09:38.596
do spatial sampling, you can also use [br]mobile sensors and Dr.

0:09:38.596,0:09:43.152
Unnikrishnan here, a post doc in the lab, [br]has worked on this or generalization of

0:09:43.152,0:09:47.708
the theory of sampling when you have [br]mobile sensors that can actually go over

0:09:47.708,0:09:56.360
a field in an arbitrary fashion. [br]Then you maybe show this in an example.

0:09:56.360,0:10:00.203
It's again a temperature monitoring [br]example here on the EPFL campus, or you

0:10:00.203,0:10:03.971
have buildings. [br]You have that open space between

0:10:03.971,0:10:07.176
buildings. [br]Those buildings are, of course, hot.

0:10:07.176,0:10:11.327
The open space are cool. [br]And you would like to have monitoring of

0:10:11.327,0:10:16.699
this temperature field not with static [br]spatial sensors, but with people running

0:10:16.699,0:10:23.960
around, having a thermal meter let's say [br]on their mobile phone.

0:10:23.960,0:10:27.854
And the question is, how accurate can you [br]actually measure temperature using a

0:10:27.854,0:10:32.240
device like this? [br]And so, this is being done actually for

0:10:32.240,0:10:37.612
pollution monitoring in the city of [br]Lausanne so there's some equipment put on

0:10:37.612,0:10:44.827
buses to measure pollution parameters. [br]And what we do here is we try to develop

0:10:44.827,0:10:49.909
a theory of how good you can sample when [br]you have these mobile sensors going over

0:10:49.909,0:10:57.310
a surface and measuring a field. [br]The results are very mathematical but are

0:10:57.310,0:11:00.660
interesting because our non-trivial [br]extension of sampling theory through

0:11:00.660,0:11:04.614
multiple dimensions. [br]And a few papers are mentioned here if

0:11:04.614,0:11:11.010
you are interested in more detail. [br]The next project is about a new way of

0:11:11.010,0:11:15.470
doing image acquisition. [br]So in this class, we have seen sampling

0:11:15.470,0:11:20.484
and we have seen quantization. [br]And when we do quantization typically we

0:11:20.484,0:11:25.460
say, let's take [UNKNOWN] samples and [br]then take as many bits as possible.

0:11:25.460,0:11:30.920
Let's say eight bits for speech, 12 bits [br]for images 24 bits maybe for audio,

0:11:30.920,0:11:35.310
etcetera. [br]Now here we took the extreme other

0:11:35.310,0:11:40.332
example we said lets build an image [br]sensor that has many, many, many pixels

0:11:40.332,0:11:46.690
but the pixels only detect either a [br]enough light or not.

0:11:46.690,0:11:49.080
So the pixels are actually binary [br]detectors.

0:11:49.080,0:11:54.610
And so you have a light intensity here. [br]Which changes over space.

0:11:54.610,0:11:58.200
You have a lens that smooths the light [br]intensity.

0:11:58.200,0:12:01.110
So what reaches the camera is this smooth [br]curve here.

0:12:01.110,0:12:05.060
And this smooth curve you sample very, [br]very, very finely.

0:12:05.060,0:12:09.410
But you only decide if it's above or [br]below a certain threshold.

0:12:09.410,0:12:13.480
So the sensor only generates a sequence [br]of binary digits.

0:12:13.480,0:12:17.850
So that's the imaging model. [br]And this has been studied by Dr.

0:12:17.850,0:12:22.008
Feng Yang, did his PhD thesis on this, is [br]now a post-doc working on this project,

0:12:22.008,0:12:27.370
and a whole slew of other people. [br]This was a very extensive project.

0:12:27.370,0:12:32.765
And what is interesting is that this new [br]way of acquiring images, for example,

0:12:32.765,0:12:40.089
allows to do high dynamic range imaging. [br]Here is a simulation of a high dynamic

0:12:40.089,0:12:45.870
range image in a much easier way than [br]with conventional cameras.

0:12:45.870,0:12:49.736
That's one advantage. [br]Another one is that you can have very,

0:12:49.736,0:12:52.615
very cheap sensors. [br]So here's an example of one that was

0:12:52.615,0:12:56.800
built in the lab. [br]And then, you take many, many frames.

0:12:56.800,0:12:59.832
They are extremely noisy. [br]If they look noisy, they are simply

0:12:59.832,0:13:03.660
binary, so you only have zeroes and ones, [br]but you have enough of these, and you do

0:13:03.660,0:13:09.955
an optimal reconstruction method. [br]You actually can recognize here, the logo

0:13:09.955,0:13:15.026
of EPFL. [br]There are publications here that you are

0:13:15.026,0:13:19.380
welcome to look up. [br]And the thesis is online.

0:13:19.380,0:13:23.230
Last but not least Rambus silicon valley [br]company, actually works with us on this

0:13:23.230,0:13:28.620
and has acquired some of the technologies [br]that was developed in this project.

0:13:30.790,0:13:33.600
And old classic is trying to predict the [br]stock market.

0:13:33.600,0:13:38.438
So, we gave it another shot. [br]so Lionel Coulot did his PhD thesis, was

0:13:38.438,0:13:43.100
co-advised with Peter Bossaerts who is at [br]Caltech.

0:13:43.100,0:13:47.316
And we were trying to understand if [br]methods from information theory would

0:13:47.316,0:13:51.668
allow to predict models for the stock [br]market, and that requires statistical

0:13:51.668,0:13:56.234
models for what the stock market might [br]be.

0:13:56.234,0:14:00.264
And what is interesting is that you have [br]to decide between very sophisticated

0:14:00.264,0:14:04.294
models that might be overkill and are [br]hard to estimate, and very simple models

0:14:04.294,0:14:08.200
which might be too simplistic, but which [br]might be very robust to things that

0:14:08.200,0:14:15.118
happen in the stock market. [br]And, in the end we used coding theory and

0:14:15.118,0:14:20.082
classic algorithmic methods like dynamic [br]programming to come up with a method that

0:14:20.082,0:14:24.434
decides what is the correct model at [br]every time of, the observation of the

0:14:24.434,0:14:31.690
stock market. [br]So I'm just going to show a picture.

0:14:31.690,0:14:35.939
And the picture is, is a value on the [br]stock market.

0:14:35.939,0:14:40.412
And the question is, can you detect if [br]the stock market is in a bear market or a

0:14:40.412,0:14:44.562
bull market? [br]So when the stock market goes up it's,

0:14:44.562,0:14:48.130
called bull market. [br]If it goes down, it's a bear market.

0:14:48.130,0:14:53.180
What is very hard is to decide by [br]watching every day what's happening.

0:14:53.180,0:14:56.582
If currently the trend is going up or the [br]trend is going down and you need to do

0:14:56.582,0:15:01.857
this with an online algorithm. [br]Okay, you cannot look into the future and

0:15:01.857,0:15:06.543
this method developed by Lionel allows to [br]do a model fitting and to very quickly

0:15:06.543,0:15:14.050
detect when the stock market changes from [br]a bull market to a bar, bear market.

0:15:15.710,0:15:21.970
The thesis online and this was sponsored [br]by, as you may guess, by a bank.

0:15:21.970,0:15:26.041
And the results are interesting, but we [br]are still having a regular day job so you

0:15:26.041,0:15:29.404
can guess that the method is not [br]completely fool proof to predict the

0:15:29.404,0:15:33.590
stock market. [br]But the methods, the algorithms, and the

0:15:33.590,0:15:39.296
theory behind it is quite cool. [br]The next few projects are so called

0:15:39.296,0:15:42.708
inverse problems. [br]So inverse problems are problems where

0:15:42.708,0:15:46.236
you have some measurements but the [br]measurements do not describe the signal

0:15:46.236,0:15:50.649
you're interested in. [br]But some indirect measurement of the

0:15:50.649,0:15:56.210
signal, so you try to invert the system [br]to go back to the original signal.

0:15:56.210,0:15:59.801
You all know about computerized [br]tomography, a medical image method, where

0:15:59.801,0:16:03.660
you can see inside the body without [br]really going there.

0:16:03.660,0:16:08.004
And that's a typical inverse problem. [br]Here we are interested in inverse

0:16:08.004,0:16:13.110
problems in environmental monitoring. [br]So, the first example is diffusion

0:16:13.110,0:16:16.545
equation. [br]And we have a physical phenomena, for

0:16:16.545,0:16:22.150
example temperature has been discussed, [br]or atmospheric dispersal of pollution.

0:16:22.150,0:16:26.840
We want to measure the field at locations [br]where we can put sensors, and the goal is

0:16:26.840,0:16:32.110
to find where are the sources, for [br]example, of pollution.

0:16:32.110,0:16:36.867
Now this is a hard problem because, you [br]have to model how, for example, pollution

0:16:36.867,0:16:40.867
is being diffused. [br]That depends on weather patterns and so

0:16:40.867,0:16:44.037
on. [br]But the tools we are using are typical

0:16:44.037,0:16:48.945
signal processing tools, for analysis. [br]Sampling theory for exemplifying finite

0:16:48.945,0:16:52.245
rate of innovation sampling or [br]compressive sensing, that has also been

0:16:52.245,0:16:56.860
mentioned earlier. [br]Let's look at the picture.

0:16:56.860,0:17:01.488
That's a very simple example of this. [br]Assume you have two smokestacks and

0:17:01.488,0:17:06.992
inside a factory compound, and the [br]smokestacks produce pollution which

0:17:06.992,0:17:12.210
changes every day. [br]You don't know how much pollution is

0:17:12.210,0:17:16.820
being released, and you're working for an [br]environmental monitoring agency.

0:17:16.820,0:17:21.644
You put sensors outside of the compound [br]and you measure what arrives, in terms of

0:17:21.644,0:17:26.569
pollution, at these sensors. [br]And the goal is to figure out if what

0:17:26.569,0:17:29.985
came out of smoke stack was within the [br]bounds allowed, lets say by z,

0:17:29.985,0:17:35.966
Environmental Protection Agency. [br]So this is a interesting and non-trivial

0:17:35.966,0:17:40.776
problem but there are some interesting [br]results that were produced by Yuri

0:17:40.776,0:17:46.104
Ranieri, whom you all know because he was [br]the famous Master Chief assistant for the

0:17:46.104,0:17:53.992
BSB class. [br]So we are able to recover sparse sources

0:17:53.992,0:17:58.378
using this inversion method. [br]we use this finite rate of innovation

0:17:58.378,0:18:02.630
sampling techniques to actually do it. [br]And here we is a list of publications

0:18:02.630,0:18:10.180
that came out of this research. [br]This problem is also an inverse problem.

0:18:10.180,0:18:13.740
It's a Fukushima inverse problem. [br]It is a PhD project of Marta

0:18:13.740,0:18:19.152
Martinez-Camara, and a few other of us [br]are involved in this, and we collaborate

0:18:19.152,0:18:26.180
with a specialist Andreas Stohl. [br]Who is a specialist of monitoring of

0:18:26.180,0:18:30.610
radioactive diffusion. [br]So what we like to do is figure out how

0:18:30.610,0:18:36.280
much radionuclides were actually released [br]in Fukushima at the time of the of the

0:18:36.280,0:18:42.560
nuclear accident at Fukushima. [br]We have only very few sensors, they are

0:18:42.560,0:18:46.060
located around the world very far away [br]from Fukushima.

0:18:46.060,0:18:51.085
And the question is, is it possible from [br]these few measurements around the world

0:18:51.085,0:18:55.885
taken later, to invert the entire process [br]as I diffuse the initial release of

0:18:55.885,0:19:03.150
radioactive material into the atmosphere. [br]What tools are we using?

0:19:03.150,0:19:06.300
Sparse regularizations, so that's [br]compressed sensing.

0:19:06.300,0:19:09.663
And we need to using atmospheric [br]dispersion model to understand how

0:19:09.663,0:19:13.616
radioactive material from from Fukushima [br]was actually transported across the

0:19:13.616,0:19:19.568
world. [br]So one result that we have and which is

0:19:19.568,0:19:24.260
very interesting is we were able to [br]estimate the emission of Xenons, that's

0:19:24.260,0:19:28.884
radioactive gas that was released at the [br]time of explosions at Fukushima, went up

0:19:28.884,0:19:36.798
into the atmosphere, was transported by [br]weather patterns all over the world.

0:19:36.798,0:19:41.218
And from the measurements all over the [br]world, we were able to pinpoint exactly

0:19:41.218,0:19:48.560
when the Xenon was released, and how much [br]Xenon was released into the atmosphere.

0:19:48.560,0:19:52.272
And it turns out we actually know the [br]total amount of Xenon that was released,

0:19:52.272,0:19:57.960
because after the accident no Xenon was [br]actually left in the nuclear power plant.

0:19:59.890,0:20:03.754
Currently we're trying to go beyond this [br]and estimate the Cesium release, but that

0:20:03.754,0:20:08.518
turns out to be a harder problem. [br]The paper that describes this will be

0:20:08.518,0:20:13.860
published ICASSP this year and is [br]available online here in infoscience.

0:20:16.390,0:20:20.614
Last but not least is a project we call, [br]"Can One Hear the Shape of a Room?" It's

0:20:20.614,0:20:24.574
a PhD project of Ivan Dokmanic and [br]several other people in the lab, in

0:20:24.574,0:20:31.110
particular, Reza Parhizkar, Andreaz [br]Walther, have worked on this.

0:20:31.110,0:20:33.806
And also we have a collaboration with Yue [br]Lu.

0:20:33.806,0:20:38.650
He's now with Harvard. [br]Now you know about this problem because,

0:20:38.650,0:20:43.980
Ivan gave module 512 about gear [br]dereverberation, echo cancellation.

0:20:43.980,0:20:48.628
And, uh,the next step is to say, if I [br]listen to echoes, can I actually

0:20:48.628,0:20:55.008
understand what is a room shape? [br]So if I know the room shape, then I know

0:20:55.008,0:20:59.898
how to generate the echoes. [br]But if you give me the echoes, can I know

0:20:59.898,0:21:03.165
the room shape? [br]It's a classic inverse problem, very cute

0:21:03.165,0:21:05.981
one. [br]And we usually explain it by saying,

0:21:05.981,0:21:10.470
let's say you enter a room, you're [br]blindfolded.

0:21:10.470,0:21:13.820
And so you don't see the room at all. [br]You snap your finger.

0:21:13.820,0:21:19.280
You therefore elicit echoes, you listen [br]very carefully to the echoes.

0:21:19.280,0:21:22.460
Can you exactly see or hear the shape of [br]the room?

0:21:24.390,0:21:27.756
Now this has a beautiful theory, which we [br]won't have time to really explain, but

0:21:27.756,0:21:31.660
that you can read up about because it's [br]published material.

0:21:31.660,0:21:36.490
But if you have a source or receiver you [br]have a direct pass between the source and

0:21:36.490,0:21:41.540
the receiver, and you have echoes given [br]by the walls.

0:21:41.540,0:21:45.895
The echoes given by the walls correspond [br]to so called mirror or image sources, so

0:21:45.895,0:21:49.665
this is the same as if you had a source [br]here and the sound would have gone

0:21:49.665,0:21:54.666
straight here. [br]So if you can locate all these image

0:21:54.666,0:21:58.750
sources, then you can actually locate the [br]room.

0:21:58.750,0:22:02.772
The walls, therefore the room. [br]And this is, you know, in principal

0:22:02.772,0:22:08.590
do-able the question was is it always [br]true that this can be done?

0:22:08.590,0:22:11.501
And is it also realistic to do it in [br]practice?

0:22:11.501,0:22:15.659
So, here are examples of a system with [br]five microforms, you have one source five

0:22:15.659,0:22:19.466
microforms. [br]You have somebody snap his finger and you

0:22:19.466,0:22:23.498
have the echos related to the walls and [br]you see there is a complexity which is,

0:22:23.498,0:22:27.467
these echos come in random orders because [br]different walls are at different

0:22:27.467,0:22:33.798
distances of the microphone. [br]And the question is, can we find out the

0:22:33.798,0:22:37.780
shape from a set of measurements as we [br]see here?

0:22:37.780,0:22:44.050
How many measurements do we need? [br]Can we have a robust algorithm?

0:22:44.050,0:22:48.490
So the answer is summarized in, yes we [br]can.

0:22:48.490,0:22:52.157
And there are some experiments we did, [br]both at the labs.

0:22:52.157,0:22:56.122
So this is one of our seminar rooms we [br]created a, an artificial wall here to

0:22:56.122,0:23:01.330
have different shapes of rooms. [br]So this is a typical shape of room.

0:23:01.330,0:23:06.082
Then in this case, with five microphone [br]and one source, we were able to estimate

0:23:06.082,0:23:12.290
the size, shape of the room very [br]accurately to more, better than 1%.

0:23:12.290,0:23:16.996
And once we had this, we said, well, [br]let's see how robust this is.

0:23:16.996,0:23:21.284
We went to Lausanne Cathedral and that's [br]actually a foyer of the Lausanne

0:23:21.284,0:23:25.840
Cathedral, which is not at all needing [br]the assumptions of the algorithms that

0:23:25.840,0:23:31.677
I've described very briefly here. [br]And it was still possible to see the

0:23:31.677,0:23:36.420
major refractors, meaning the major walls [br]here in the Lausanne Cathedral.

0:23:36.420,0:23:40.200
And so the answer is yes, one can hear [br]the shape of a room.

0:23:40.200,0:23:44.840
And you can visit Ivan's web page to see [br]more details.

0:23:46.980,0:23:50.518
Now these were just a selection of [br]projects, of works that is being done by

0:23:50.518,0:23:55.100
PhDs and post-docs and senior researchers [br]in the lab.

0:23:55.100,0:23:59.126
Please go to the website, as that gives [br]the entire portfolio of research here of

0:23:59.126,0:24:02.010
what the lab is currently doing.