[Script Info]
Title: 
[Events]
Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text
Dialogue: 0,0:00:00.66,0:00:03.96,Default,,0000,0000,0000,,Module 10.2. \NSome research projects that use the
Dialogue: 0,0:00:03.96,0:00:07.77,Default,,0000,0000,0000,,techniques you have learned in the \Ndigital signal processing class.
Dialogue: 0,0:00:09.05,0:00:11.89,Default,,0000,0000,0000,,We're going to talk about some current \Nresearch in the lab.
Dialogue: 0,0:00:11.89,0:00:15.98,Default,,0000,0000,0000,,There is a whole slew of them, and it's a \Nselection here of interesting research
Dialogue: 0,0:00:15.98,0:00:21.64,Default,,0000,0000,0000,,projects that we can briefly discuss. \NThe first one is, eFacsimile is a project
Dialogue: 0,0:00:21.64,0:00:26.67,Default,,0000,0000,0000,,on art work acquisition. \NThe second one is about signal processing
Dialogue: 0,0:00:26.67,0:00:30.17,Default,,0000,0000,0000,,in sensor networks. \NThen there is a network science result on
Dialogue: 0,0:00:30.17,0:00:35.02,Default,,0000,0000,0000,,source localization put in graphs. \NThen we talk about sampling result, so
Dialogue: 0,0:00:35.02,0:00:38.69,Default,,0000,0000,0000,,called finite rate of innovation \Nsampling.
Dialogue: 0,0:00:38.69,0:00:42.47,Default,,0000,0000,0000,,Then we talk again about sampling, that \Nof physical fields, using some new
Dialogue: 0,0:00:42.47,0:00:47.26,Default,,0000,0000,0000,,techniques for sampling. \NThen, we have a project on image
Dialogue: 0,0:00:47.26,0:00:52.72,Default,,0000,0000,0000,,acquisition where we change the sensors \Nused in acquiring images.
Dialogue: 0,0:00:52.72,0:00:55.27,Default,,0000,0000,0000,,Then, an old classic, predicting the \Nstock market.
Dialogue: 0,0:00:56.55,0:01:00.59,Default,,0000,0000,0000,,Then, we talk about inverse problems. \NThe three next projects are actually
Dialogue: 0,0:01:00.59,0:01:04.21,Default,,0000,0000,0000,,inverse problems. \NThe first one is on the diffusion
Dialogue: 0,0:01:04.21,0:01:08.36,Default,,0000,0000,0000,,equation, the second one is trying to \Nunderstand the nuclear fall out from
Dialogue: 0,0:01:08.36,0:01:14.40,Default,,0000,0000,0000,,Fukushima, and last but not least is an \Ninverse problem in acoustics.
Dialogue: 0,0:01:17.46,0:01:20.48,Default,,0000,0000,0000,,The eFacsimile project. \NThis is a project that we do together
Dialogue: 0,0:01:20.48,0:01:26.80,Default,,0000,0000,0000,,with Google to try to improve how artwork \Nis represented on the internet.
Dialogue: 0,0:01:26.80,0:01:31.91,Default,,0000,0000,0000,,it's lead by [INAUDIBLE] researcher Loic \NBaboulaz and several PhD students are
Dialogue: 0,0:01:31.91,0:01:36.12,Default,,0000,0000,0000,,involved in this. \NThe questions are how to capture,
Dialogue: 0,0:01:36.12,0:01:40.67,Default,,0000,0000,0000,,represent, and render artwork as well as \Npossible.
Dialogue: 0,0:01:40.67,0:01:45.42,Default,,0000,0000,0000,,And to do this, we need some advanced \Ntechniques on relighting, manipulation
Dialogue: 0,0:01:45.42,0:01:50.45,Default,,0000,0000,0000,,of, the so called, light fields that is \Nacquired, and potentially high resolution
Dialogue: 0,0:01:50.45,0:01:57.54,Default,,0000,0000,0000,,solutions for mobile devices. \NThere are some demos online that I
Dialogue: 0,0:01:57.54,0:02:01.66,Default,,0000,0000,0000,,encourage you to actually watch. \Nbecause this really doesn't show the
Dialogue: 0,0:02:01.66,0:02:04.62,Default,,0000,0000,0000,,idea. \NThis is of course a static version.
Dialogue: 0,0:02:04.62,0:02:08.93,Default,,0000,0000,0000,,But for example one of the demos is you \Ntake a, an oil painting here.
Dialogue: 0,0:02:08.93,0:02:13.21,Default,,0000,0000,0000,,And you acquire it in such a way that if \Nyou show it on a tablet and you move it,
Dialogue: 0,0:02:13.21,0:02:19.14,Default,,0000,0000,0000,,it will actually exactly look like the \Noriginal oil painting.
Dialogue: 0,0:02:19.14,0:02:22.45,Default,,0000,0000,0000,,So you get the illusion that you have \Nactually the oil painting in the hand.
Dialogue: 0,0:02:22.45,0:02:26.55,Default,,0000,0000,0000,,So if the light changes, the vis, \Nvisualization will change.
Dialogue: 0,0:02:26.55,0:02:29.43,Default,,0000,0000,0000,,If you turn the tablet, the visualization \Nwill change.
Dialogue: 0,0:02:29.43,0:02:34.33,Default,,0000,0000,0000,,so then, what is quite stunning and I \Nsuggest you actually watch it, similarly,
Dialogue: 0,0:02:34.33,0:02:39.39,Default,,0000,0000,0000,,there is another demo which deals with \Nstain glasses.
Dialogue: 0,0:02:39.39,0:02:42.91,Default,,0000,0000,0000,,So stain glasses are very interesting art \Nobjects, but very difficult to render on
Dialogue: 0,0:02:42.91,0:02:46.28,Default,,0000,0000,0000,,the internet. \NAnd so here we will stimulate the stained
Dialogue: 0,0:02:46.28,0:02:49.76,Default,,0000,0000,0000,,glass, so if you have a tablet in your \Nhand and you move it, it looks like if
Dialogue: 0,0:02:49.76,0:02:56.77,Default,,0000,0000,0000,,you had, stained glass in your hand. \NSo the tools we use is, we use
Dialogue: 0,0:02:56.77,0:03:01.75,Default,,0000,0000,0000,,traditional cameras, but we also use \Nso-called light field cameras.
Dialogue: 0,0:03:01.75,0:03:03.93,Default,,0000,0000,0000,,You might have heard of the light \N[INAUDIBLE] for example.
Dialogue: 0,0:03:03.93,0:03:07.17,Default,,0000,0000,0000,,That's a new generation of camera. \NIt's extremely interesting.
Dialogue: 0,0:03:07.17,0:03:11.88,Default,,0000,0000,0000,,And so we need to fully understand light \Ntransport theory and [INAUDIBLE].
Dialogue: 0,0:03:11.88,0:03:15.42,Default,,0000,0000,0000,,Which uses sparse recovery methods or \Ncompress sensing.
Dialogue: 0,0:03:15.42,0:03:20.80,Default,,0000,0000,0000,,The website of the project is given here. \NAnd as I indicated there is YouTube demo
Dialogue: 0,0:03:20.80,0:03:25.55,Default,,0000,0000,0000,,that shows quite realistically the demos \Nthat we're discussed just in a minute
Dialogue: 0,0:03:25.55,0:03:31.42,Default,,0000,0000,0000,,ago. \NSo, next project is about wireless sensor
Dialogue: 0,0:03:31.42,0:03:36.51,Default,,0000,0000,0000,,networks, in particular about monitoring \Nvisually in a wireless sensor network.
Dialogue: 0,0:03:36.51,0:03:39.23,Default,,0000,0000,0000,,So, sensor networks have deployed for \Nmany years.
Dialogue: 0,0:03:39.23,0:03:43.86,Default,,0000,0000,0000,,We have large projects here, in the lab, \Non environmental monitoring.
Dialogue: 0,0:03:43.86,0:03:48.01,Default,,0000,0000,0000,,And the current generation is actually \Nequipped with camera, and then you have a
Dialogue: 0,0:03:48.01,0:03:51.88,Default,,0000,0000,0000,,problem of compression, of \Nrepresentation.
Dialogue: 0,0:03:51.88,0:03:55.48,Default,,0000,0000,0000,,So even though the trend is towards \Nsmaller and smaller devices, they are
Dialogue: 0,0:03:55.48,0:03:59.26,Default,,0000,0000,0000,,still power hungry and in particular if \Nyou have a sophisticated camera, the
Dialogue: 0,0:03:59.26,0:04:03.33,Default,,0000,0000,0000,,number of images or the number of pixels \Nthat is generated might actually overwelm
Dialogue: 0,0:04:03.33,0:04:08.98,Default,,0000,0000,0000,,the power budget of the system. \NAnd so Dr.
Dialogue: 0,0:04:08.98,0:04:13.14,Default,,0000,0000,0000,,Zichong Chen, who finished his PhD here, \Nand did his post doc, together with
Dialogue: 0,0:04:13.14,0:04:17.43,Default,,0000,0000,0000,,Guillermo Barrenetxea, are looking at \Ncreating large scale sensor networks
Dialogue: 0,0:04:17.43,0:04:22.29,Default,,0000,0000,0000,,equipped with cameras that are energy \Nefficient.
Dialogue: 0,0:04:24.05,0:04:28.62,Default,,0000,0000,0000,,So why do we want images? \NWell, here are a few examples.
Dialogue: 0,0:04:28.62,0:04:33.25,Default,,0000,0000,0000,,This is from I think a Berkeley project \Nabout, monitoring birds' nest.
Dialogue: 0,0:04:33.25,0:04:36.89,Default,,0000,0000,0000,,Unfortunately a snake is showing up an \Nhe's actually eating all the eggs in the
Dialogue: 0,0:04:36.89,0:04:40.38,Default,,0000,0000,0000,,nest. \NSo that's, monitoring for why life
Dialogue: 0,0:04:40.38,0:04:43.46,Default,,0000,0000,0000,,protection. \NHere is an example from the Swiss Alps
Dialogue: 0,0:04:43.46,0:04:48.12,Default,,0000,0000,0000,,monitoring for for avalanche detection. \NHere is also an example from the Swiss
Dialogue: 0,0:04:48.12,0:04:51.85,Default,,0000,0000,0000,,Alps, its monitoring to see weather \Nconditions.
Dialogue: 0,0:04:51.85,0:04:55.81,Default,,0000,0000,0000,,And, finally, here is monitoring networks \Nthat is installed on the PFL campus.
Dialogue: 0,0:04:55.81,0:05:00.99,Default,,0000,0000,0000,,In all these cases you have many cameras \Nusing small communication devices.
Dialogue: 0,0:05:00.99,0:05:04.69,Default,,0000,0000,0000,,And so compression and representation is \Nextremely critical.
Dialogue: 0,0:05:06.63,0:05:12.62,Default,,0000,0000,0000,,So there are a number of results that you \Ncan find in the thesis of Dr.
Dialogue: 0,0:05:12.62,0:05:17.44,Default,,0000,0000,0000,,Chi Chong Chang, given here in this, \Nwebsite, and essentially the idea is
Dialogue: 0,0:05:17.44,0:05:21.96,Default,,0000,0000,0000,,that, cameras can help each other to \Nreduce the amount of information that
Dialogue: 0,0:05:21.96,0:05:26.20,Default,,0000,0000,0000,,actually has to be sent to the base \Nstation or into the cloud, for doing
Dialogue: 0,0:05:26.20,0:05:32.87,Default,,0000,0000,0000,,efficient monitoring. \NSo a long with signal processing today,
Dialogue: 0,0:05:32.87,0:05:36.28,Default,,0000,0000,0000,,actually it's moving to single processing \Non graphs.
Dialogue: 0,0:05:36.28,0:05:39.82,Default,,0000,0000,0000,,I don't have to explain to you the \Nimportance, for example, of social
Dialogue: 0,0:05:39.82,0:05:43.51,Default,,0000,0000,0000,,networks. \NAnd so Pedro Pinto who was involved here
Dialogue: 0,0:05:43.51,0:05:47.61,Default,,0000,0000,0000,,in the class and is a post doc in the \Nlab, together with Patrick Thiran, has
Dialogue: 0,0:05:47.61,0:05:52.23,Default,,0000,0000,0000,,worked on the problem of source \Nlocalization.
Dialogue: 0,0:05:52.23,0:05:57.16,Default,,0000,0000,0000,,So you have some graph here, let's say \Nsocial network and somebody launches a
Dialogue: 0,0:05:57.16,0:06:00.85,Default,,0000,0000,0000,,rumor. \NHere is the source and the rumor gets
Dialogue: 0,0:06:00.85,0:06:05.25,Default,,0000,0000,0000,,forwarded along the edges of the graph at \Ndifferent times and you have some
Dialogue: 0,0:06:05.25,0:06:12.57,Default,,0000,0000,0000,,observers, say green nodes here, that \Nreceives a rumor at some instant of time.
Dialogue: 0,0:06:12.57,0:06:15.77,Default,,0000,0000,0000,,They know where the rumor comes from, you \Nknow who told you the gossip, and you
Dialogue: 0,0:06:15.77,0:06:19.98,Default,,0000,0000,0000,,know when you got the gossip information. \NSo, the question is, you know the
Dialogue: 0,0:06:19.98,0:06:23.00,Default,,0000,0000,0000,,structure of the graph, or you have an \Napproximation of the structure of the
Dialogue: 0,0:06:23.00,0:06:26.38,Default,,0000,0000,0000,,graphs. \NYou have these observations.
Dialogue: 0,0:06:26.38,0:06:30.29,Default,,0000,0000,0000,,Can you figure out who actually, spreads \Nthe rumor first.
Dialogue: 0,0:06:30.29,0:06:33.45,Default,,0000,0000,0000,,It turns out this has, an interesting \Nsolution.
Dialogue: 0,0:06:33.45,0:06:38.35,Default,,0000,0000,0000,,And using only few observers, about 20%, \Nyou can achieve a very high accuracy in
Dialogue: 0,0:06:38.35,0:06:43.51,Default,,0000,0000,0000,,finding the source of a rumor on a large \Nscale network.
Dialogue: 0,0:06:43.51,0:06:46.93,Default,,0000,0000,0000,,And there are many interesting questions \Nhere, to pursue in this source
Dialogue: 0,0:06:46.93,0:06:51.29,Default,,0000,0000,0000,,localization in social networks. \NAnd there was a paper that came out last
Dialogue: 0,0:06:51.29,0:06:55.45,Default,,0000,0000,0000,,year, Locating the Source of Diffusion in \NLarge-Scale Networks that had quite a bit
Dialogue: 0,0:06:55.45,0:06:59.10,Default,,0000,0000,0000,,of impact. \NThe project is actually funded by the
Dialogue: 0,0:06:59.10,0:07:02.60,Default,,0000,0000,0000,,Bill and Melinda Gates Foundation. \NThe reason is that one of the
Dialogue: 0,0:07:02.60,0:07:05.65,Default,,0000,0000,0000,,applications is to monitor health \Nproblems.
Dialogue: 0,0:07:05.65,0:07:10.64,Default,,0000,0000,0000,,For example, here is a map of Cholera \Noutbreak in Africa, and the map shows the
Dialogue: 0,0:07:10.64,0:07:15.37,Default,,0000,0000,0000,,river network. \NCholera is a water born disease, and so,
Dialogue: 0,0:07:15.37,0:07:19.74,Default,,0000,0000,0000,,typically Cholera will actually diffuse \Nalong waterways.
Dialogue: 0,0:07:19.74,0:07:23.77,Default,,0000,0000,0000,,But you know when people fell sick at \Ncertain locations and then you can infer
Dialogue: 0,0:07:23.77,0:07:27.41,Default,,0000,0000,0000,,the source of the actual Cholera \Noutbreak.
Dialogue: 0,0:07:27.41,0:07:31.37,Default,,0000,0000,0000,,There is another example here, which is a \Nsimulation of, if you had to figure out
Dialogue: 0,0:07:31.37,0:07:35.15,Default,,0000,0000,0000,,if there was some pollution or attack on \Nthe New York subway, and if you could
Dialogue: 0,0:07:35.15,0:07:39.23,Default,,0000,0000,0000,,figure out knowing the network of the New \NYork subway and when you start detecting
Dialogue: 0,0:07:39.23,0:07:45.75,Default,,0000,0000,0000,,the problems where the source of the \Nproblem actually was.
Dialogue: 0,0:07:47.31,0:07:51.28,Default,,0000,0000,0000,,The next project is on sampling, so we \Nhave worked on a new theory of sampling
Dialogue: 0,0:07:51.28,0:07:54.68,Default,,0000,0000,0000,,here called Finite Rate of Innovation \NSampling, and it is used in
Dialogue: 0,0:07:54.68,0:07:58.65,Default,,0000,0000,0000,,communications problems, and in \Nmonitoring problems to reduce the number
Dialogue: 0,0:07:58.65,0:08:03.96,Default,,0000,0000,0000,,of samples being transmitted or acquired. \NDr.
Dialogue: 0,0:08:03.96,0:08:08.45,Default,,0000,0000,0000,,Freris, who is a senior scientist with \Ndoctoral students and MS assistants, are
Dialogue: 0,0:08:08.45,0:08:12.87,Default,,0000,0000,0000,,actually working on doing ECG monitoring \Nat very low power for wireless health
Dialogue: 0,0:08:12.87,0:08:17.00,Default,,0000,0000,0000,,monitoring. \NSo here is a block diagram, it's
Dialogue: 0,0:08:17.00,0:08:21.03,Default,,0000,0000,0000,,relatively complicated so let me not get \Ninto this, but it uses some fairly
Dialogue: 0,0:08:21.03,0:08:25.19,Default,,0000,0000,0000,,sophisticated techniques to reduce the \Nsampling rate so as to reduce the energy
Dialogue: 0,0:08:25.19,0:08:33.08,Default,,0000,0000,0000,,consumption on these wireless devices. \NSo, this project is actually sponsored by
Dialogue: 0,0:08:33.08,0:08:40.16,Default,,0000,0000,0000,,somebody well known, Qualcomm, interested \Nin the theory of sampling.
Dialogue: 0,0:08:40.16,0:08:43.64,Default,,0000,0000,0000,,And the extension here for this \Nparticular project has been
Dialogue: 0,0:08:43.64,0:08:49.81,Default,,0000,0000,0000,,generalization of the initial finite rate \Nof innovation sampling methodology.
Dialogue: 0,0:08:49.81,0:08:53.73,Default,,0000,0000,0000,,To get better compression, and better \Nmodelization of the signals.
Dialogue: 0,0:08:53.73,0:08:57.82,Default,,0000,0000,0000,,So here we have the ECG signal, and then \Nthere is sophisticated models that
Dialogue: 0,0:08:57.82,0:09:02.80,Default,,0000,0000,0000,,allows, to take very few parameters, to \Nmodel the ECG signal.
Dialogue: 0,0:09:02.80,0:09:06.50,Default,,0000,0000,0000,,There are a number of papers here, the \Ninitial paper on finite rate of
Dialogue: 0,0:09:06.50,0:09:10.92,Default,,0000,0000,0000,,innovation sampling is this 2002 paper, \Nand the number of recent papers have done
Dialogue: 0,0:09:10.92,0:09:16.94,Default,,0000,0000,0000,,extension to this theory. \NSo if you like sampling I welcome you to
Dialogue: 0,0:09:16.94,0:09:22.75,Default,,0000,0000,0000,,actually read up on this stuff, it's one \Nof my favorite research topics.
Dialogue: 0,0:09:24.88,0:09:28.84,Default,,0000,0000,0000,,When we talk about sampling already in \Nsensor networks we have mentioned that
Dialogue: 0,0:09:28.84,0:09:34.36,Default,,0000,0000,0000,,placing a sensor is like taking a sample. \NAnd so that spatial sampling, now if you
Dialogue: 0,0:09:34.36,0:09:38.60,Default,,0000,0000,0000,,do spatial sampling, you can also use \Nmobile sensors and Dr.
Dialogue: 0,0:09:38.60,0:09:43.15,Default,,0000,0000,0000,,Unnikrishnan here, a post doc in the lab, \Nhas worked on this or generalization of
Dialogue: 0,0:09:43.15,0:09:47.71,Default,,0000,0000,0000,,the theory of sampling when you have \Nmobile sensors that can actually go over
Dialogue: 0,0:09:47.71,0:09:56.36,Default,,0000,0000,0000,,a field in an arbitrary fashion. \NThen you maybe show this in an example.
Dialogue: 0,0:09:56.36,0:10:00.20,Default,,0000,0000,0000,,It's again a temperature monitoring \Nexample here on the EPFL campus, or you
Dialogue: 0,0:10:00.20,0:10:03.97,Default,,0000,0000,0000,,have buildings. \NYou have that open space between
Dialogue: 0,0:10:03.97,0:10:07.18,Default,,0000,0000,0000,,buildings. \NThose buildings are, of course, hot.
Dialogue: 0,0:10:07.18,0:10:11.33,Default,,0000,0000,0000,,The open space are cool. \NAnd you would like to have monitoring of
Dialogue: 0,0:10:11.33,0:10:16.70,Default,,0000,0000,0000,,this temperature field not with static \Nspatial sensors, but with people running
Dialogue: 0,0:10:16.70,0:10:23.96,Default,,0000,0000,0000,,around, having a thermal meter let's say \Non their mobile phone.
Dialogue: 0,0:10:23.96,0:10:27.85,Default,,0000,0000,0000,,And the question is, how accurate can you \Nactually measure temperature using a
Dialogue: 0,0:10:27.85,0:10:32.24,Default,,0000,0000,0000,,device like this? \NAnd so, this is being done actually for
Dialogue: 0,0:10:32.24,0:10:37.61,Default,,0000,0000,0000,,pollution monitoring in the city of \NLausanne so there's some equipment put on
Dialogue: 0,0:10:37.61,0:10:44.83,Default,,0000,0000,0000,,buses to measure pollution parameters. \NAnd what we do here is we try to develop
Dialogue: 0,0:10:44.83,0:10:49.91,Default,,0000,0000,0000,,a theory of how good you can sample when \Nyou have these mobile sensors going over
Dialogue: 0,0:10:49.91,0:10:57.31,Default,,0000,0000,0000,,a surface and measuring a field. \NThe results are very mathematical but are
Dialogue: 0,0:10:57.31,0:11:00.66,Default,,0000,0000,0000,,interesting because our non-trivial \Nextension of sampling theory through
Dialogue: 0,0:11:00.66,0:11:04.61,Default,,0000,0000,0000,,multiple dimensions. \NAnd a few papers are mentioned here if
Dialogue: 0,0:11:04.61,0:11:11.01,Default,,0000,0000,0000,,you are interested in more detail. \NThe next project is about a new way of
Dialogue: 0,0:11:11.01,0:11:15.47,Default,,0000,0000,0000,,doing image acquisition. \NSo in this class, we have seen sampling
Dialogue: 0,0:11:15.47,0:11:20.48,Default,,0000,0000,0000,,and we have seen quantization. \NAnd when we do quantization typically we
Dialogue: 0,0:11:20.48,0:11:25.46,Default,,0000,0000,0000,,say, let's take [UNKNOWN] samples and \Nthen take as many bits as possible.
Dialogue: 0,0:11:25.46,0:11:30.92,Default,,0000,0000,0000,,Let's say eight bits for speech, 12 bits \Nfor images 24 bits maybe for audio,
Dialogue: 0,0:11:30.92,0:11:35.31,Default,,0000,0000,0000,,etcetera. \NNow here we took the extreme other
Dialogue: 0,0:11:35.31,0:11:40.33,Default,,0000,0000,0000,,example we said lets build an image \Nsensor that has many, many, many pixels
Dialogue: 0,0:11:40.33,0:11:46.69,Default,,0000,0000,0000,,but the pixels only detect either a \Nenough light or not.
Dialogue: 0,0:11:46.69,0:11:49.08,Default,,0000,0000,0000,,So the pixels are actually binary \Ndetectors.
Dialogue: 0,0:11:49.08,0:11:54.61,Default,,0000,0000,0000,,And so you have a light intensity here. \NWhich changes over space.
Dialogue: 0,0:11:54.61,0:11:58.20,Default,,0000,0000,0000,,You have a lens that smooths the light \Nintensity.
Dialogue: 0,0:11:58.20,0:12:01.11,Default,,0000,0000,0000,,So what reaches the camera is this smooth \Ncurve here.
Dialogue: 0,0:12:01.11,0:12:05.06,Default,,0000,0000,0000,,And this smooth curve you sample very, \Nvery, very finely.
Dialogue: 0,0:12:05.06,0:12:09.41,Default,,0000,0000,0000,,But you only decide if it's above or \Nbelow a certain threshold.
Dialogue: 0,0:12:09.41,0:12:13.48,Default,,0000,0000,0000,,So the sensor only generates a sequence \Nof binary digits.
Dialogue: 0,0:12:13.48,0:12:17.85,Default,,0000,0000,0000,,So that's the imaging model. \NAnd this has been studied by Dr.
Dialogue: 0,0:12:17.85,0:12:22.01,Default,,0000,0000,0000,,Feng Yang, did his PhD thesis on this, is \Nnow a post-doc working on this project,
Dialogue: 0,0:12:22.01,0:12:27.37,Default,,0000,0000,0000,,and a whole slew of other people. \NThis was a very extensive project.
Dialogue: 0,0:12:27.37,0:12:32.76,Default,,0000,0000,0000,,And what is interesting is that this new \Nway of acquiring images, for example,
Dialogue: 0,0:12:32.76,0:12:40.09,Default,,0000,0000,0000,,allows to do high dynamic range imaging. \NHere is a simulation of a high dynamic
Dialogue: 0,0:12:40.09,0:12:45.87,Default,,0000,0000,0000,,range image in a much easier way than \Nwith conventional cameras.
Dialogue: 0,0:12:45.87,0:12:49.74,Default,,0000,0000,0000,,That's one advantage. \NAnother one is that you can have very,
Dialogue: 0,0:12:49.74,0:12:52.62,Default,,0000,0000,0000,,very cheap sensors. \NSo here's an example of one that was
Dialogue: 0,0:12:52.62,0:12:56.80,Default,,0000,0000,0000,,built in the lab. \NAnd then, you take many, many frames.
Dialogue: 0,0:12:56.80,0:12:59.83,Default,,0000,0000,0000,,They are extremely noisy. \NIf they look noisy, they are simply
Dialogue: 0,0:12:59.83,0:13:03.66,Default,,0000,0000,0000,,binary, so you only have zeroes and ones, \Nbut you have enough of these, and you do
Dialogue: 0,0:13:03.66,0:13:09.96,Default,,0000,0000,0000,,an optimal reconstruction method. \NYou actually can recognize here, the logo
Dialogue: 0,0:13:09.96,0:13:15.03,Default,,0000,0000,0000,,of EPFL. \NThere are publications here that you are
Dialogue: 0,0:13:15.03,0:13:19.38,Default,,0000,0000,0000,,welcome to look up. \NAnd the thesis is online.
Dialogue: 0,0:13:19.38,0:13:23.23,Default,,0000,0000,0000,,Last but not least Rambus silicon valley \Ncompany, actually works with us on this
Dialogue: 0,0:13:23.23,0:13:28.62,Default,,0000,0000,0000,,and has acquired some of the technologies \Nthat was developed in this project.
Dialogue: 0,0:13:30.79,0:13:33.60,Default,,0000,0000,0000,,And old classic is trying to predict the \Nstock market.
Dialogue: 0,0:13:33.60,0:13:38.44,Default,,0000,0000,0000,,So, we gave it another shot. \Nso Lionel Coulot did his PhD thesis, was
Dialogue: 0,0:13:38.44,0:13:43.10,Default,,0000,0000,0000,,co-advised with Peter Bossaerts who is at \NCaltech.
Dialogue: 0,0:13:43.10,0:13:47.32,Default,,0000,0000,0000,,And we were trying to understand if \Nmethods from information theory would
Dialogue: 0,0:13:47.32,0:13:51.67,Default,,0000,0000,0000,,allow to predict models for the stock \Nmarket, and that requires statistical
Dialogue: 0,0:13:51.67,0:13:56.23,Default,,0000,0000,0000,,models for what the stock market might \Nbe.
Dialogue: 0,0:13:56.23,0:14:00.26,Default,,0000,0000,0000,,And what is interesting is that you have \Nto decide between very sophisticated
Dialogue: 0,0:14:00.26,0:14:04.29,Default,,0000,0000,0000,,models that might be overkill and are \Nhard to estimate, and very simple models
Dialogue: 0,0:14:04.29,0:14:08.20,Default,,0000,0000,0000,,which might be too simplistic, but which \Nmight be very robust to things that
Dialogue: 0,0:14:08.20,0:14:15.12,Default,,0000,0000,0000,,happen in the stock market. \NAnd, in the end we used coding theory and
Dialogue: 0,0:14:15.12,0:14:20.08,Default,,0000,0000,0000,,classic algorithmic methods like dynamic \Nprogramming to come up with a method that
Dialogue: 0,0:14:20.08,0:14:24.43,Default,,0000,0000,0000,,decides what is the correct model at \Nevery time of, the observation of the
Dialogue: 0,0:14:24.43,0:14:31.69,Default,,0000,0000,0000,,stock market. \NSo I'm just going to show a picture.
Dialogue: 0,0:14:31.69,0:14:35.94,Default,,0000,0000,0000,,And the picture is, is a value on the \Nstock market.
Dialogue: 0,0:14:35.94,0:14:40.41,Default,,0000,0000,0000,,And the question is, can you detect if \Nthe stock market is in a bear market or a
Dialogue: 0,0:14:40.41,0:14:44.56,Default,,0000,0000,0000,,bull market? \NSo when the stock market goes up it's,
Dialogue: 0,0:14:44.56,0:14:48.13,Default,,0000,0000,0000,,called bull market. \NIf it goes down, it's a bear market.
Dialogue: 0,0:14:48.13,0:14:53.18,Default,,0000,0000,0000,,What is very hard is to decide by \Nwatching every day what's happening.
Dialogue: 0,0:14:53.18,0:14:56.58,Default,,0000,0000,0000,,If currently the trend is going up or the \Ntrend is going down and you need to do
Dialogue: 0,0:14:56.58,0:15:01.86,Default,,0000,0000,0000,,this with an online algorithm. \NOkay, you cannot look into the future and
Dialogue: 0,0:15:01.86,0:15:06.54,Default,,0000,0000,0000,,this method developed by Lionel allows to \Ndo a model fitting and to very quickly
Dialogue: 0,0:15:06.54,0:15:14.05,Default,,0000,0000,0000,,detect when the stock market changes from \Na bull market to a bar, bear market.
Dialogue: 0,0:15:15.71,0:15:21.97,Default,,0000,0000,0000,,The thesis online and this was sponsored \Nby, as you may guess, by a bank.
Dialogue: 0,0:15:21.97,0:15:26.04,Default,,0000,0000,0000,,And the results are interesting, but we \Nare still having a regular day job so you
Dialogue: 0,0:15:26.04,0:15:29.40,Default,,0000,0000,0000,,can guess that the method is not \Ncompletely fool proof to predict the
Dialogue: 0,0:15:29.40,0:15:33.59,Default,,0000,0000,0000,,stock market. \NBut the methods, the algorithms, and the
Dialogue: 0,0:15:33.59,0:15:39.30,Default,,0000,0000,0000,,theory behind it is quite cool. \NThe next few projects are so called
Dialogue: 0,0:15:39.30,0:15:42.71,Default,,0000,0000,0000,,inverse problems. \NSo inverse problems are problems where
Dialogue: 0,0:15:42.71,0:15:46.24,Default,,0000,0000,0000,,you have some measurements but the \Nmeasurements do not describe the signal
Dialogue: 0,0:15:46.24,0:15:50.65,Default,,0000,0000,0000,,you're interested in. \NBut some indirect measurement of the
Dialogue: 0,0:15:50.65,0:15:56.21,Default,,0000,0000,0000,,signal, so you try to invert the system \Nto go back to the original signal.
Dialogue: 0,0:15:56.21,0:15:59.80,Default,,0000,0000,0000,,You all know about computerized \Ntomography, a medical image method, where
Dialogue: 0,0:15:59.80,0:16:03.66,Default,,0000,0000,0000,,you can see inside the body without \Nreally going there.
Dialogue: 0,0:16:03.66,0:16:08.00,Default,,0000,0000,0000,,And that's a typical inverse problem. \NHere we are interested in inverse
Dialogue: 0,0:16:08.00,0:16:13.11,Default,,0000,0000,0000,,problems in environmental monitoring. \NSo, the first example is diffusion
Dialogue: 0,0:16:13.11,0:16:16.54,Default,,0000,0000,0000,,equation. \NAnd we have a physical phenomena, for
Dialogue: 0,0:16:16.54,0:16:22.15,Default,,0000,0000,0000,,example temperature has been discussed, \Nor atmospheric dispersal of pollution.
Dialogue: 0,0:16:22.15,0:16:26.84,Default,,0000,0000,0000,,We want to measure the field at locations \Nwhere we can put sensors, and the goal is
Dialogue: 0,0:16:26.84,0:16:32.11,Default,,0000,0000,0000,,to find where are the sources, for \Nexample, of pollution.
Dialogue: 0,0:16:32.11,0:16:36.87,Default,,0000,0000,0000,,Now this is a hard problem because, you \Nhave to model how, for example, pollution
Dialogue: 0,0:16:36.87,0:16:40.87,Default,,0000,0000,0000,,is being diffused. \NThat depends on weather patterns and so
Dialogue: 0,0:16:40.87,0:16:44.04,Default,,0000,0000,0000,,on. \NBut the tools we are using are typical
Dialogue: 0,0:16:44.04,0:16:48.94,Default,,0000,0000,0000,,signal processing tools, for analysis. \NSampling theory for exemplifying finite
Dialogue: 0,0:16:48.94,0:16:52.24,Default,,0000,0000,0000,,rate of innovation sampling or \Ncompressive sensing, that has also been
Dialogue: 0,0:16:52.24,0:16:56.86,Default,,0000,0000,0000,,mentioned earlier. \NLet's look at the picture.
Dialogue: 0,0:16:56.86,0:17:01.49,Default,,0000,0000,0000,,That's a very simple example of this. \NAssume you have two smokestacks and
Dialogue: 0,0:17:01.49,0:17:06.99,Default,,0000,0000,0000,,inside a factory compound, and the \Nsmokestacks produce pollution which
Dialogue: 0,0:17:06.99,0:17:12.21,Default,,0000,0000,0000,,changes every day. \NYou don't know how much pollution is
Dialogue: 0,0:17:12.21,0:17:16.82,Default,,0000,0000,0000,,being released, and you're working for an \Nenvironmental monitoring agency.
Dialogue: 0,0:17:16.82,0:17:21.64,Default,,0000,0000,0000,,You put sensors outside of the compound \Nand you measure what arrives, in terms of
Dialogue: 0,0:17:21.64,0:17:26.57,Default,,0000,0000,0000,,pollution, at these sensors. \NAnd the goal is to figure out if what
Dialogue: 0,0:17:26.57,0:17:29.98,Default,,0000,0000,0000,,came out of smoke stack was within the \Nbounds allowed, lets say by z,
Dialogue: 0,0:17:29.98,0:17:35.97,Default,,0000,0000,0000,,Environmental Protection Agency. \NSo this is a interesting and non-trivial
Dialogue: 0,0:17:35.97,0:17:40.78,Default,,0000,0000,0000,,problem but there are some interesting \Nresults that were produced by Yuri
Dialogue: 0,0:17:40.78,0:17:46.10,Default,,0000,0000,0000,,Ranieri, whom you all know because he was \Nthe famous Master Chief assistant for the
Dialogue: 0,0:17:46.10,0:17:53.99,Default,,0000,0000,0000,,BSB class. \NSo we are able to recover sparse sources
Dialogue: 0,0:17:53.99,0:17:58.38,Default,,0000,0000,0000,,using this inversion method. \Nwe use this finite rate of innovation
Dialogue: 0,0:17:58.38,0:18:02.63,Default,,0000,0000,0000,,sampling techniques to actually do it. \NAnd here we is a list of publications
Dialogue: 0,0:18:02.63,0:18:10.18,Default,,0000,0000,0000,,that came out of this research. \NThis problem is also an inverse problem.
Dialogue: 0,0:18:10.18,0:18:13.74,Default,,0000,0000,0000,,It's a Fukushima inverse problem. \NIt is a PhD project of Marta
Dialogue: 0,0:18:13.74,0:18:19.15,Default,,0000,0000,0000,,Martinez-Camara, and a few other of us \Nare involved in this, and we collaborate
Dialogue: 0,0:18:19.15,0:18:26.18,Default,,0000,0000,0000,,with a specialist Andreas Stohl. \NWho is a specialist of monitoring of
Dialogue: 0,0:18:26.18,0:18:30.61,Default,,0000,0000,0000,,radioactive diffusion. \NSo what we like to do is figure out how
Dialogue: 0,0:18:30.61,0:18:36.28,Default,,0000,0000,0000,,much radionuclides were actually released \Nin Fukushima at the time of the of the
Dialogue: 0,0:18:36.28,0:18:42.56,Default,,0000,0000,0000,,nuclear accident at Fukushima. \NWe have only very few sensors, they are
Dialogue: 0,0:18:42.56,0:18:46.06,Default,,0000,0000,0000,,located around the world very far away \Nfrom Fukushima.
Dialogue: 0,0:18:46.06,0:18:51.08,Default,,0000,0000,0000,,And the question is, is it possible from \Nthese few measurements around the world
Dialogue: 0,0:18:51.08,0:18:55.88,Default,,0000,0000,0000,,taken later, to invert the entire process \Nas I diffuse the initial release of
Dialogue: 0,0:18:55.88,0:19:03.15,Default,,0000,0000,0000,,radioactive material into the atmosphere. \NWhat tools are we using?
Dialogue: 0,0:19:03.15,0:19:06.30,Default,,0000,0000,0000,,Sparse regularizations, so that's \Ncompressed sensing.
Dialogue: 0,0:19:06.30,0:19:09.66,Default,,0000,0000,0000,,And we need to using atmospheric \Ndispersion model to understand how
Dialogue: 0,0:19:09.66,0:19:13.62,Default,,0000,0000,0000,,radioactive material from from Fukushima \Nwas actually transported across the
Dialogue: 0,0:19:13.62,0:19:19.57,Default,,0000,0000,0000,,world. \NSo one result that we have and which is
Dialogue: 0,0:19:19.57,0:19:24.26,Default,,0000,0000,0000,,very interesting is we were able to \Nestimate the emission of Xenons, that's
Dialogue: 0,0:19:24.26,0:19:28.88,Default,,0000,0000,0000,,radioactive gas that was released at the \Ntime of explosions at Fukushima, went up
Dialogue: 0,0:19:28.88,0:19:36.80,Default,,0000,0000,0000,,into the atmosphere, was transported by \Nweather patterns all over the world.
Dialogue: 0,0:19:36.80,0:19:41.22,Default,,0000,0000,0000,,And from the measurements all over the \Nworld, we were able to pinpoint exactly
Dialogue: 0,0:19:41.22,0:19:48.56,Default,,0000,0000,0000,,when the Xenon was released, and how much \NXenon was released into the atmosphere.
Dialogue: 0,0:19:48.56,0:19:52.27,Default,,0000,0000,0000,,And it turns out we actually know the \Ntotal amount of Xenon that was released,
Dialogue: 0,0:19:52.27,0:19:57.96,Default,,0000,0000,0000,,because after the accident no Xenon was \Nactually left in the nuclear power plant.
Dialogue: 0,0:19:59.89,0:20:03.75,Default,,0000,0000,0000,,Currently we're trying to go beyond this \Nand estimate the Cesium release, but that
Dialogue: 0,0:20:03.75,0:20:08.52,Default,,0000,0000,0000,,turns out to be a harder problem. \NThe paper that describes this will be
Dialogue: 0,0:20:08.52,0:20:13.86,Default,,0000,0000,0000,,published ICASSP this year and is \Navailable online here in infoscience.
Dialogue: 0,0:20:16.39,0:20:20.61,Default,,0000,0000,0000,,Last but not least is a project we call, \N"Can One Hear the Shape of a Room?" It's
Dialogue: 0,0:20:20.61,0:20:24.57,Default,,0000,0000,0000,,a PhD project of Ivan Dokmanic and \Nseveral other people in the lab, in
Dialogue: 0,0:20:24.57,0:20:31.11,Default,,0000,0000,0000,,particular, Reza Parhizkar, Andreaz \NWalther, have worked on this.
Dialogue: 0,0:20:31.11,0:20:33.81,Default,,0000,0000,0000,,And also we have a collaboration with Yue \NLu.
Dialogue: 0,0:20:33.81,0:20:38.65,Default,,0000,0000,0000,,He's now with Harvard. \NNow you know about this problem because,
Dialogue: 0,0:20:38.65,0:20:43.98,Default,,0000,0000,0000,,Ivan gave module 512 about gear \Ndereverberation, echo cancellation.
Dialogue: 0,0:20:43.98,0:20:48.63,Default,,0000,0000,0000,,And, uh,the next step is to say, if I \Nlisten to echoes, can I actually
Dialogue: 0,0:20:48.63,0:20:55.01,Default,,0000,0000,0000,,understand what is a room shape? \NSo if I know the room shape, then I know
Dialogue: 0,0:20:55.01,0:20:59.90,Default,,0000,0000,0000,,how to generate the echoes. \NBut if you give me the echoes, can I know
Dialogue: 0,0:20:59.90,0:21:03.16,Default,,0000,0000,0000,,the room shape? \NIt's a classic inverse problem, very cute
Dialogue: 0,0:21:03.16,0:21:05.98,Default,,0000,0000,0000,,one. \NAnd we usually explain it by saying,
Dialogue: 0,0:21:05.98,0:21:10.47,Default,,0000,0000,0000,,let's say you enter a room, you're \Nblindfolded.
Dialogue: 0,0:21:10.47,0:21:13.82,Default,,0000,0000,0000,,And so you don't see the room at all. \NYou snap your finger.
Dialogue: 0,0:21:13.82,0:21:19.28,Default,,0000,0000,0000,,You therefore elicit echoes, you listen \Nvery carefully to the echoes.
Dialogue: 0,0:21:19.28,0:21:22.46,Default,,0000,0000,0000,,Can you exactly see or hear the shape of \Nthe room?
Dialogue: 0,0:21:24.39,0:21:27.76,Default,,0000,0000,0000,,Now this has a beautiful theory, which we \Nwon't have time to really explain, but
Dialogue: 0,0:21:27.76,0:21:31.66,Default,,0000,0000,0000,,that you can read up about because it's \Npublished material.
Dialogue: 0,0:21:31.66,0:21:36.49,Default,,0000,0000,0000,,But if you have a source or receiver you \Nhave a direct pass between the source and
Dialogue: 0,0:21:36.49,0:21:41.54,Default,,0000,0000,0000,,the receiver, and you have echoes given \Nby the walls.
Dialogue: 0,0:21:41.54,0:21:45.90,Default,,0000,0000,0000,,The echoes given by the walls correspond \Nto so called mirror or image sources, so
Dialogue: 0,0:21:45.90,0:21:49.66,Default,,0000,0000,0000,,this is the same as if you had a source \Nhere and the sound would have gone
Dialogue: 0,0:21:49.66,0:21:54.67,Default,,0000,0000,0000,,straight here. \NSo if you can locate all these image
Dialogue: 0,0:21:54.67,0:21:58.75,Default,,0000,0000,0000,,sources, then you can actually locate the \Nroom.
Dialogue: 0,0:21:58.75,0:22:02.77,Default,,0000,0000,0000,,The walls, therefore the room. \NAnd this is, you know, in principal
Dialogue: 0,0:22:02.77,0:22:08.59,Default,,0000,0000,0000,,do-able the question was is it always \Ntrue that this can be done?
Dialogue: 0,0:22:08.59,0:22:11.50,Default,,0000,0000,0000,,And is it also realistic to do it in \Npractice?
Dialogue: 0,0:22:11.50,0:22:15.66,Default,,0000,0000,0000,,So, here are examples of a system with \Nfive microforms, you have one source five
Dialogue: 0,0:22:15.66,0:22:19.47,Default,,0000,0000,0000,,microforms. \NYou have somebody snap his finger and you
Dialogue: 0,0:22:19.47,0:22:23.50,Default,,0000,0000,0000,,have the echos related to the walls and \Nyou see there is a complexity which is,
Dialogue: 0,0:22:23.50,0:22:27.47,Default,,0000,0000,0000,,these echos come in random orders because \Ndifferent walls are at different
Dialogue: 0,0:22:27.47,0:22:33.80,Default,,0000,0000,0000,,distances of the microphone. \NAnd the question is, can we find out the
Dialogue: 0,0:22:33.80,0:22:37.78,Default,,0000,0000,0000,,shape from a set of measurements as we \Nsee here?
Dialogue: 0,0:22:37.78,0:22:44.05,Default,,0000,0000,0000,,How many measurements do we need? \NCan we have a robust algorithm?
Dialogue: 0,0:22:44.05,0:22:48.49,Default,,0000,0000,0000,,So the answer is summarized in, yes we \Ncan.
Dialogue: 0,0:22:48.49,0:22:52.16,Default,,0000,0000,0000,,And there are some experiments we did, \Nboth at the labs.
Dialogue: 0,0:22:52.16,0:22:56.12,Default,,0000,0000,0000,,So this is one of our seminar rooms we \Ncreated a, an artificial wall here to
Dialogue: 0,0:22:56.12,0:23:01.33,Default,,0000,0000,0000,,have different shapes of rooms. \NSo this is a typical shape of room.
Dialogue: 0,0:23:01.33,0:23:06.08,Default,,0000,0000,0000,,Then in this case, with five microphone \Nand one source, we were able to estimate
Dialogue: 0,0:23:06.08,0:23:12.29,Default,,0000,0000,0000,,the size, shape of the room very \Naccurately to more, better than 1%.
Dialogue: 0,0:23:12.29,0:23:16.100,Default,,0000,0000,0000,,And once we had this, we said, well, \Nlet's see how robust this is.
Dialogue: 0,0:23:16.100,0:23:21.28,Default,,0000,0000,0000,,We went to Lausanne Cathedral and that's \Nactually a foyer of the Lausanne
Dialogue: 0,0:23:21.28,0:23:25.84,Default,,0000,0000,0000,,Cathedral, which is not at all needing \Nthe assumptions of the algorithms that
Dialogue: 0,0:23:25.84,0:23:31.68,Default,,0000,0000,0000,,I've described very briefly here. \NAnd it was still possible to see the
Dialogue: 0,0:23:31.68,0:23:36.42,Default,,0000,0000,0000,,major refractors, meaning the major walls \Nhere in the Lausanne Cathedral.
Dialogue: 0,0:23:36.42,0:23:40.20,Default,,0000,0000,0000,,And so the answer is yes, one can hear \Nthe shape of a room.
Dialogue: 0,0:23:40.20,0:23:44.84,Default,,0000,0000,0000,,And you can visit Ivan's web page to see \Nmore details.
Dialogue: 0,0:23:46.98,0:23:50.52,Default,,0000,0000,0000,,Now these were just a selection of \Nprojects, of works that is being done by
Dialogue: 0,0:23:50.52,0:23:55.10,Default,,0000,0000,0000,,PhDs and post-docs and senior researchers \Nin the lab.
Dialogue: 0,0:23:55.10,0:23:59.13,Default,,0000,0000,0000,,Please go to the website, as that gives \Nthe entire portfolio of research here of
Dialogue: 0,0:23:59.13,0:24:02.01,Default,,0000,0000,0000,,what the lab is currently doing.