Player is loading...

Embed

Embed code

Transcriptions

Note: this content has been automatically generated.
00:00:00
which for the introduction i'm welcome everybody to the second session so
00:00:04
um can machine earning help you discover what you don't know
00:00:08
um so let me first give an overview of the group so the uncertainty qualification and optimal design group was
00:00:16
created in step that on september two thousand fifteen and so this is soon the third anniversary and um
00:00:23
we are currently you a very small group there is one permanent one intern and one visit
00:00:28
to master students but they already a six a and nine uh including visiting students
00:00:34
and uh as i haven't mentioned awesome special economic ties to a university of them
00:00:39
and in particular the institute of not directed at medical statistics and uh actuarial sciences
00:00:45
since the last uh advisory board um there's been one projects that was
00:00:50
completed in the group which is learning and is very easy
00:00:54
usually saying a dangerous regions in most right part arm into space which was funded by the uh has information
00:01:01
i would comment on the outcomes of this project in a in a few slides
00:01:06
and uh now we have a new project that is starting um was two
00:01:10
pages students that are hired uh in the frame of the project
00:01:14
uncertainty qualification and efficient design of experiments for data and simulation for even
00:01:19
inverse problem solving that is funded by this with national science foundation
00:01:25
so um it's about small data indeed and um
00:01:30
the group focuses on production optimisation and inversion
00:01:34
uh in the framework of complex systems that are typically quite expensive to sample up to evaluate
00:01:41
and the core methodological approaches of the group uh rely on a gaussian process modelling
00:01:48
a basin optimisation base inset estimation and and sequential
00:01:52
design algorithms named sequential uncertainty reduction artworks
00:01:57
so the main application domains and partners include the
00:02:01
environmental joe sciences with uh several collaborations with
00:02:05
union was an initial then uh the french geological survey b. r. d. or g. m.
00:02:12
accept rock uh energy engineering with the french nuclear safety institute's also with
00:02:18
the kind of your those salt dodge energetic and see that nothing
00:02:22
and also uh inclined dynamics was the uh national centre for climate change research and um
00:02:29
also this is a group that has a reparations with other groups so uh we started the uh some
00:02:35
preparations in a robotic spectacle busted bar and the watching by metrics uh what's the best around them
00:02:43
so uh the kind of uh the flagship example that some of you might already
00:02:47
know of of the group is uh um uh for all the problem
00:02:52
so here we are looking at a a simulation of how a contaminant
00:02:57
care and propagate in a subset it's so it's a two d.
00:03:02
section of a of a phonetically fair so a a natural uh has a lot of water
00:03:10
and we are uh assuming that at time zero there
00:03:13
is the contaminants that is starting its uh
00:03:19
flow in the subset face and here we are monitoring at different wells
00:03:26
so you can see the triangles though this pen for wells
00:03:29
was the colours that are paired with the this uh the two
00:03:32
figures where we're observing the concentration of contaminants all the time
00:03:37
so at every well you have this uh the the flow is
00:03:42
taking place we measure the consultation contaminants and then the game
00:03:46
is uh to try to recover the starting position of the contaminants from discussed
00:03:54
which seems like a mission impossible because you have the the information you have
00:03:59
is is relatively little compared to the complexity of ah of the phenomenon
00:04:05
but we are trying to use a base in experimental design and basin of
00:04:09
summarisation algorithms to to to recover this so here this is complete
00:04:14
a synthetic best case where everything's controlled so we know whether component was started
00:04:19
because uh it was imposed the simulation and we are relying on the on the steps of salvation you're you're trying to recover this
00:04:26
so we'll come back to the sinner insect but before that a bit well that's that that's uh let me come back
00:04:39
a notion of uncertainty quantification so uh you might that's not well uh be uh
00:04:45
informed yet about what you q. means ostensible so it's a relatively recent dissipate
00:04:51
it's a discipline that is intrinsically uh at the interface between
00:04:55
several disciplines which or name the scientific computing and statistics
00:05:01
and the goal of you q. is to predict complex systems um
00:05:05
like a pink numerical simulations and and data driven statistical procedures
00:05:10
so um this field as has really been from rushing
00:05:15
in recent years in particular are there is a
00:05:17
total that was announced into doesn't thirteen by uh that's i am and a essay that all that
00:05:24
a society of industrial not by mathematics and the uh american statistical association
00:05:28
uh and uh and also the there was recently a six month program
00:05:33
at the isaac newton's you didn't enrich about you q. for
00:05:36
complex systems period of the budget so this is just to show
00:05:39
you a two two examples of this uh drawing uh
00:05:45
field and ah optimal design in the name of the group stands
00:05:51
primarily for model based design of experiments so in particular experiments at the very expensive
00:05:56
uh to perform a with an emphasis on sequential design of numerical simulations
00:06:03
design of experiments is a diary of statistics and design of numerical simulation is an area that
00:06:09
is developing um because more and more we replace
00:06:13
a year field experiments by artificial experience
00:06:18
so if we come back to our uh for all problem here you have an example
00:06:22
of how this base unit to my session and reasons the work correctly so
00:06:28
here we are looking at the objective function the objective function you if you
00:06:33
remember that that's case you had these three different carols that you wanted
00:06:39
a base on in order to involve the system and find the contaminant
00:06:44
location so basically if you take any comedic location for the contaminants
00:06:49
source you can run the full simulation also the candles and calculated the misfits between the
00:06:56
reference cows and the hypothetical uh curves obtained
00:07:00
by moving the contaminate contamination source
00:07:04
and then if you look at this picture yeah what you see is how this
00:07:09
integrated at all is behaving when you move the hypothetical location of the contaminants
00:07:15
and what you see is that there is a unique point of minimum
00:07:17
your green and this point is actually the real actual oh
00:07:24
place where the contaminants uh start it's and what you want is to find
00:07:28
this place which is basically a global optimisation problem you want to find
00:07:31
the global minimum of this very complicated normally they are horrible curve or not
00:07:37
uh based on a few experts you start by doing so naive experiments the blue
00:07:43
points of schools based in in design and then you can approximate this misfits
00:07:51
based on the nine a blue points and you ups you get a a very very naive representation
00:07:56
of this complicated function based on mine what's your but that also comes with some uh
00:08:02
creation of production and so tainted and by combining these two maths you can
00:08:08
guy either a sequential design main meaning that you can add
00:08:12
new simulations in order to hopefully converge with this
00:08:16
a green ones so here uh we are using the so called expect improvement your
00:08:22
stick strategy to our uh uh points which is uh the right point
00:08:27
and meaning that we we're going to really perform the simulation at this point and
00:08:34
then we are going to calculate the misfits and we will see that this orange new points here
00:08:39
is affecting the model is affecting the answer change c. n. is meeting asked when you are yeah and then
00:08:45
we're going to iterate this and by iterating that we're
00:08:49
going to explore this complicated mourning out on stage
00:08:54
and eventually after about twenty rounds we're going to explore and to find the global optimum
00:09:01
get so instead of running full grades of calculation over the two dimensions here we can
00:09:08
uh rely on a relatively moderate budget and find the source of contamination of
00:09:14
course this is a two example once a small dimension there's a lot of work
00:09:18
going on to expand this uh to higher dimensional problems and many other changes
00:09:25
so this is a test case that you might have already have uh come across in particular uh if you remember for
00:09:32
the yeah people the son of g. p.'s that'd yep last year that we uh animated without your the monty
00:09:39
there is also prevent available on this test case and this is a joint work with people project can with
00:09:44
the the job i don't i'm using this boldface uh to uh the notes uh the the um alumni
00:09:51
now she's in a she's a permanent a lecture at the medical university in the bangkok thailand
00:09:57
young people is a little thawing as opposed docket inverse of the japanese also collaboration misshapen now uh the shot that
00:10:05
but that wasn't so i'm just of couple of words about recent extensions of this so we
00:10:11
have a uh just try to to to do batch sequential design on this and show
00:10:17
would the obtain some nice results uh showing the the benefits of but analysing the expiration
00:10:23
on this this case uh with this whole teen and adult amount as well and
00:10:29
we are currently considering an extension of that to uh the case where the digit lodges and something as well
00:10:37
so now let me um speak briefly about some recent achievements
00:10:43
first about the as a project on a learning in the visualising dinners regions
00:10:48
and then um the very short overview of some other recent contributions in finite finally
00:10:54
i if time allows i uh speak briefly about some projects ongoing out it's
00:11:00
so this has the project has enabled the support of that yachtsman see as a post op
00:11:05
ah let me just mention that shall have been before long term visit thoughts you uh and the group while a
00:11:11
performing is is p. h. d. studies that you don't do it now as opposed look at it yeah
00:11:17
and this is a project that has a involved a a more a test
00:11:22
case with the french geological survey that that we'll speak about right now
00:11:28
so they'll simulating a frauds and assimilating
00:11:31
basically a coastal flooding risks um
00:11:36
yeah the atlantic shore in uh in france yeah that was set and uh they're particularly focusing here
00:11:42
on the small regions around the site people showed up and the they are looking at
00:11:47
um the um the prodding expand its depending on
00:11:53
the wave and other physical uh parameters
00:11:57
so the operator rising the simulation using five inputs um that all the
00:12:03
forcing parameters that or formant rising that fight fight and search and
00:12:07
the or more the link the uh the fraud in impact over the
00:12:11
in aunt ah and in particular the total are we offering
00:12:16
as a function of this five arm is what they want to know is what is the regions
00:12:20
of the five dimensional space uh that all leading to a floating above a given fresh
00:12:28
so uh we is that it would not go to get the l. g. m. with developed the method based on uh
00:12:33
courtenay profile max of um a profile maxima quite tonight of annexing off a
00:12:40
that is a a method to visualise an expression sets in high dimensions
00:12:45
based on the maximum of or the remaining viable so for
00:12:49
each viable weekend brought how the supreme i'm off
00:12:53
the the complex system is behaving when this dimension is is change
00:12:59
it is a very complicated function because this is a supreme uh so it's uh
00:13:03
it's really it's missus states a lot of uh evaluations and since we cannot
00:13:09
perform a lot of evaluation we use gaussian process is but if you take the supreme of goes imposes is you get something that is not gets in
00:13:15
so did it lead us to some um relatively hardcore developments and we could come up
00:13:20
with some quantification of uncertainties on the production of the supreme of goes imposes is
00:13:25
uh which uh enable us to put you discover sent to know what's or
00:13:29
the ranges of parameters that never lead always lead to some excursions
00:13:36
so this is a developed in the paper uh that is to to be resubmitted your
00:13:40
um with the belgian collect colleagues and we have also within this project
00:13:44
this more as a project we have also uh revise the
00:13:47
paper about often probability estimation that is not published in a just
00:13:51
a g. s. and uh we have also uh oh
00:13:56
based on this we also concede the currently i'm
00:14:00
revising this work on conservative said destination which
00:14:04
is presented in more detail in the pasta but you can see and the whole
00:14:08
so these are some of the recent papers don't have time to go into detail
00:14:11
about that but just to show you will working on in finances and crawls
00:14:15
oh also was high dimensional methods withdrawals ah and the
00:14:20
uh mixtures between yeltsin process moldings and wait let's
00:14:24
also doing some um work collaborative work on your shoe
00:14:28
for hydrogen hydrogen physics where we're basically transferring recent advances and not
00:14:33
to uh apply communities and also um the formations and be
00:14:38
riveted basic control design for functions be settled and citations
00:14:43
there is also another point o. of importance is the
00:14:46
climate than energy related contributions so very gleefully
00:14:50
we have hosted um the internship and on and on net in collaboration with the with
00:14:55
the centre for climate change research and we have worked on nonstationary time series
00:15:00
of extreme values with application kind advances is is not published yet but something ongoing
00:15:05
and the uh we have also a longstanding collaboration with
00:15:09
people working notably on the modelling of hail
00:15:13
um that's the pain is a is a is an important risk for the economy
00:15:17
and the um um we have used a possum regression model uh to uh
00:15:25
to the chance of a whole day's involved in switzerland that was but which
00:15:28
wasn't enough was very research and finally at the nice good to
00:15:32
is now doing an internship on a new collaboration with the the clan where
00:15:37
we are using global sensitivity analysis to analyse impact of market parameters
00:15:42
uh on the to the optimal energy boards so we have a function
00:15:47
that is a yeah a that is the result of us over
00:15:51
we are analysing how ah changes in parameters affect the uh best
00:15:57
response of this over all the uh optimum values of inputs
00:16:04
okay ah so finally just some perspectives i'm too much of eighteen problems uh
00:16:10
for that or um underlined yes and of project that will start soon
00:16:14
so one of them is coming from computational cosmetology so here we are working always computational
00:16:20
'cause manages from a th so this is the plan and uh they have some
00:16:27
telescope images of the universe and offering to bet on this the new house and you need a cell some constants was medical
00:16:33
constants that are uh that are not known precisely but once you
00:16:38
assume values for this possible illegal constants you can simulate
00:16:41
how this telescope image ah could look like
00:16:45
well from a physical gravitational models
00:16:49
so basically the same day the whole universal portion of the universe and they come up with is
00:16:53
the distribution of active galaxies and what they want is to find those magical parameters that much
00:16:59
who is the um the statistical properties of galaxies that they actually up so
00:17:04
and for that they need methods that of sequential and that uh will choose which simulation to run
00:17:10
uh in order to best identified a spot on those and we will uh develop some novel approaches that
00:17:16
another problem we all uh i'm facing is uh the design
00:17:22
of experiments on actual uh a complex systems from nature
00:17:28
namely volcanoes so here this is a collaboration with a unicef roseanne and the group of that's in there
00:17:34
and the the idea is that we use a basin inversion so
00:17:38
here to based on reading actually measurements we want to find
00:17:42
um the density inside of a volcano and most specifically to identify regions of past activities which correspond
00:17:48
to high density or too high grade intensity of the high coverage of them to to
00:17:52
and uh what we want to do is to design the experiments actually pretty
00:17:55
much measurements on the volcano an optimal weight on the some basic model
00:18:01
so as a final word subgroup that is developing ah it's developing research
00:18:07
group is a solid methodological backgrounds and the number of research perspectives
00:18:12
and we are very open to discussions with potential partners and it's like

Share this talk: 


Conference program

Introduction by Hervé Bourlard
BOURLARD, Hervé, Idiap Director, EPFL Full Professor
29 Aug. 2018 · 9:03 a.m.
Presentation of the «Speech & Audio Processing» research group
MAGIMAI DOSS, Mathew, Idiap Senior Researcher
29 Aug. 2018 · 9:22 a.m.
Presentation of the «Robot Learning & Interaction» research group
CALINON, Sylvain, Idiap Senior Researcher
29 Aug. 2018 · 9:43 a.m.
Presentation of the «Machine Learning» research group
FLEURET, François, Idiap Senior Researcher, EPFL Maître d'enseignement et de recherche
29 Aug. 2018 · 10:04 a.m.
Presentation of the «Uncertainty Quantification and Optimal Design» research group
GINSBOURGER, David, Idiap Senior Researcher, Bern Titular Professor
29 Aug. 2018 · 11:05 a.m.
Presentation of the «Perception and Activity Understanding» research group
ODOBEZ, Jean-Marc, Idiap Senior Researcher, EPFL Maître d'enseignement et de recherche
29 Aug. 2018 · 11:24 a.m.
Presentation of the «Computational Bioimaging» research group
LIEBLING, Michael, Idiap Senior Researcher, UC Santa Barbara Adjunct Professor
29 Aug. 2018 · 11:45 a.m.
Presentation of the «Natural Language Understanding» research group
HENDERSON, James, Idiap Senior Researcher
29 Aug. 2018 · 2:03 p.m.
Presentation of the «Biometrics Security and Privacy» research group
MARCEL, Sébastien, Idiap Senior Researcher
29 Aug. 2018 · 2:19 p.m.
Presentation of the «Biosignal Processing» research group
RABELLO DOS ANJOS, André, Idiap Researcher
29 Aug. 2018 · 2:43 p.m.
Presentation of the «Social Computing» research group
GATICA-PEREZ, Daniel, Idiap Senior Researcher, EPFL Adjunct Professor
29 Aug. 2018 · 2:59 p.m.

Recommended talks

Some Challenges in Biometrics: Facial Sketch, Altered Fingerprints & SMT
Anil K. Jain, Michigan State University
3 Sept. 2013 · 2:02 p.m.