Player is loading...
Presentation of the «Machine Learning» research group
FLEURET, François, Idiap Senior Researcher, EPFL Maître d'enseignement et de recherche
Wednesday, Aug. 29, 2018 · 10:04 a.m. · 24m 13s
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:00
okay my thing so that was introduction another um so uh uh my group is a working on machine on
00:00:07
oh maybe to summarise oppression is are both in common to all
00:00:10
two groups such yep uh him or less quantities but
00:00:14
it's kind of a a running theme that uh we're all using a lot
00:00:19
or maybe to summarise a bit what machine on the news about so fashioning takes its hoots installer statistics and
00:00:25
in regression so this is uh maybe the most simple case uh you have a bunch of points so
00:00:31
a measurement that could be a john on the another measurement which would be like
00:00:35
a drug pusher or one of the last of the quantity of interest
00:00:39
i mean it's most basic for machine learning is about regression so to find some legalities and use
00:00:44
that appointed to measure so that it on you can do a prediction on in finance
00:00:49
one from zeus kind of simple case of aggression it has divorce
00:00:55
first to a more sophisticated uh on design methods so this is
00:01:00
the new station of the type of models we had in computer vision a five six years ago
00:01:06
whereas this is a model of a bike where are you
00:01:09
explicitly tells the computer that's uh the more that
00:01:13
is composed of parts on the test to learn from a lot of images of by the bicycles
00:01:18
what's possible clyde on jose are a high range which aspect which was uh in any major
00:01:23
well recently there have been a a lot of either of the the the neural networks with
00:01:28
the so called keep running a preys on the idea of deep running is simply too
00:01:34
it's in some way to avoid this home design uh mow
00:01:38
down onto simply connect a lot of very generic when
00:01:41
used together on to form that are to try to to find forms of that as of the underlying structure
00:01:47
and this is a new station of uh where i mean they keep running is been shining she's go playing
00:01:52
where a huge it was able to reach new one
00:01:56
class a skills because a lot of data available
00:02:01
so in like this what a lot of room do not part of the committee's concentrating now which is the planning thing but
00:02:08
i i have to insist that in a lot of situations uh you still need to use mall stand
00:02:13
our machine learning techniques because you don't have enough data to to train a very nice but then
00:02:19
so my group is working on the reading such methods of uh boasts
00:02:23
a store that machine methods on order so the planning methods
00:02:26
i mean them of repetition we are most interested in computer vision so as i was saying the pledge his name was
00:02:32
computer vision running now it's mentioning on computer be on on your machine running with a lot of repetition can be television
00:02:38
i think we don't vision the change that you get very high they a very high dimension input signal when you made your video
00:02:45
when you try to extract structure from the signal so we it's it's a few cases it
00:02:49
can be fine stands straight information so in that case we try to reconstruct the shape
00:02:54
of a second f. i mean yeah in a minister now on the you you you so
00:02:58
it's more like a geometrical information location of the camera on the c. d. shape
00:03:03
it can be a moss some sort of structure are a shape
00:03:08
with a lot of code topological constraints so such as
00:03:11
uh unknowns on the x. own seen any night issues oh i can be similar semantic hero type
00:03:19
of information but again the the common thing that you get a very high dimension signal
00:03:24
when you try to extract automatically a very low dimension signal on pay out in the
00:03:30
meantime of dimension compared to the original signal which can be union said dimension
00:03:34
well that i'm going to show in the coming spicy so activities so how
00:03:39
we develop new machine learning techniques to so this kind of of changes
00:03:44
so to give a bit of context in term of size on activities so my group is
00:03:49
composed of three to six pages didn't depending with the project
00:03:53
um what ought to buy stocks i we have a heavy correlation with the proposal that we
00:03:58
have its job well that people who bridges uh so he's actually do with it
00:04:04
in general a disk you have projects and i were to make it awkward each escalator pretty cool
00:04:12
uh on there so i i have a strong interaction is e. p. f. or i could surprise pages you don't so this
00:04:17
is a total mess of of what is our chosen michael we use around a six to to pay top ten
00:04:24
in term of themes we are concentrated on a sweetie aspect
00:04:28
of machine on into this of course is computation because
00:04:30
to stick nick's ass extremely heavy constraint demanding time of competition if your little
00:04:35
bit from you know with the p. q. he on that we use
00:04:38
thousands of g. p. use for the very impressive most impressive repetitions
00:04:42
we're talking about a thousand of days of g. p. u.
00:04:45
computing to to build a model which is extremely problematic was in
00:04:49
term of cost an intimate funk on on time impact
00:04:52
we also interested in memory footprint which is so much a stop light you need to actually one the computation with madame
00:04:58
that's a very limiting factor when you you want is an animated now environment such as the phone or don't
00:05:04
oh you are so interested in a small sample learning which is classes is
00:05:08
you pose it of what is one key aspect of determining which is
00:05:12
usually you have hundreds of thousands millions tens of millions and billions of examples but in many of applied
00:05:18
up the situation you you have a few hundred examples and you have to do with this
00:05:25
uh interpretation as as was scenes must uncover television we are so strong we
00:05:30
uh we're trying to do as most of the petite yep so
00:05:34
uh well maine and the production is up an academic publications on open source code that to
00:05:38
make available to the community as every was saying so we yep as a strong uh
00:05:45
or of uh i just got on spam more
00:05:48
invested i never city so we i we have a strong uh
00:05:52
like activity in that direction we years we did the
00:05:56
do you read design systems which i could probably ministry and situations
00:06:01
also because i think we have to thank them again and again and again
00:06:04
as a funding for michael can comes mostly from the sun itself
00:06:08
just a foundation so we like those guys you know she's on the art foundation so we like
00:06:14
guys because we are blessed it's it's always a fun funding agencies which are
00:06:19
he uh on the light on the on ice so we can't insisting of the buttons
00:06:25
um okay so i'm i'm going to do straight it'd be too so themes a school exam
00:06:30
purse on the um to give a feeling of the type of things we do so
00:06:36
the first big theme use oh can we speed up the is it taking
00:06:40
so can we reduce the computation uh when we reduce models from data
00:06:44
so we have to the big axes here white is to design exact methods so we had the
00:06:50
if you uh this is for the series of papers using geometrical bounce where
00:06:55
we were able to get computation by uh using exact
00:06:59
estimate of distance doing competition on doing that
00:07:02
because this was very far from caesar last time i saw this point on though my model change a little bit yeah
00:07:07
it's no use to we compute is distance which is very expensive because i know it's too far away on
00:07:12
doing c. so it's very simple thing would be to inequality from joe mcphee wherever to speed up
00:07:17
tremendously uh some very based on that goes and especially the cleaning separate unfold as we know
00:07:24
we have also another series of uh work which are uh one going
00:07:28
from which everything going for years about using something that has
00:07:33
so typically sees uh with the type of program where we want
00:07:37
to find a system to do detection of a pedestrians
00:07:41
on the the t. v. show here that's it's very easy you
00:07:44
we have usually a small number of positive images of pakistan
00:07:49
on the among the negative example so parts of when you made which are not this yeah and the vast majority of the easy
00:07:55
to stop downtown to say this is not a pedestrian because it's you know it's a flat for it's a piece of sky
00:08:00
but uh mom's billions of position in email address on a piece and some of them starts to look like
00:08:06
you to to detect the disk and to the cloud or three ten look like it is yeah
00:08:11
so it's a key issue to be able to um some
00:08:14
perl difficult exam persons uh means billions of negative examples
00:08:19
on weekends in a way simply to look at is billions of example or
00:08:22
some just check to which one looks like a and the problematic case
00:08:26
but what was this what we design where the statistics mates that's that's to get methods
00:08:31
actually very similar to the ones used to some per uh when you play
00:08:34
go essays to solve that whooping systems are based on something mechanisms
00:08:39
to concentrate computation overs appointed parton only what we see that the the doings
00:08:44
of trainings at that point was rejected on sundays of competition well
00:08:48
was concentrating trying to find negative examples problematic negative examples induce more complicated oh yeah
00:08:54
so the idea that the computer on the fly is concentrating competition on part of the data
00:08:59
which are problematic on the avoiding two two ways competition rights can be trivially aged
00:09:04
or sensitive example that you stages use um yeah we're trying to
00:09:09
train a it's wichita example where the signals to dimension
00:09:12
what do you want to change the system to makes uh to cassie fees a point above the curve
00:09:17
just as a point bureaus okay so to say all the points uh about this uh
00:09:21
kind of sine curve is one class on the point below those of gas
00:09:25
and the united way would just sample the onset points in the square on
00:09:29
say okay this is gas was asked to discuss one discuss too
00:09:33
what yeah what we see that what uh what goes on i wasn't loud goes any sampling the data set during training
00:09:39
we see that very quickly it rejects loose are are so the the
00:09:44
white part means not something a lot like part in something about
00:09:47
what we see that effectively stumping pop only a usable nine because this is what matter very quickly
00:09:53
knows the cases i know this i know this i know only start to focus the computation
00:09:58
over the boundary where actually so the problem lies it's where you want to get ones
00:10:04
what they what we see that an uh effect in actually case of handwritten digit recognition
00:10:09
do some pairs uses so it's over that the set of sixty thousand some powers
00:10:13
we see that was digits was the one some perl more than one thousand times more than he sat in sump and a new for me
00:10:19
on which makes sense because this is of these are one this digit was labelled as
00:10:24
a fall into that asset which make the system freaks not that it did
00:10:28
this is the eight so we see that actually does a good job with
00:10:31
very quickly rejected easy cases on consumer computation of the the good ones
00:10:37
so so so want aspect is no to speed up doing testing info on so
00:10:41
now it's is a model has been trained how can you do a in
00:10:44
for instance the time applications or running applications on a on a cell phone
00:10:48
so he what we proposed it's uh it was a very stoned uh
00:10:53
action algorithm which was massive used it in the computer vision domain so
00:10:58
the horrible part but then what what we propose wes when you want
00:11:01
to sort of altruism you outings as a a very similar plus
00:11:05
the uh one 'cause of competition again and again and again it's some
00:11:08
single contribution in singapore ceasing to based on down the nation
00:11:12
and it's a very well known resulting signal processing which is that we
00:11:16
can cut down computation by using so called fast with also
00:11:20
on what was interesting that even those something rude you learn in first year of university uh
00:11:25
when you still using opposing it has never been used to speed up this thing
00:11:29
on when we look at full custody that that's your brain okay we should use password control
00:11:34
but when you start to look a bit into it should be like that there are no
00:11:37
implementation details which are quite at the because now even so mathematically is very keen
00:11:42
when you want to implement it in your c. p. u. you have a lot of
00:11:46
issues related to memory could paint on the caches age question is on on the
00:11:50
technical aspect on there was a a very uh lucky to have at this time a
00:11:54
student which was spending a very large spectrum of a skins from mathematics too
00:11:59
oh see gone level programming but he was able to to we were able
00:12:04
to design a way of controlling the memory footprint on in the
00:12:07
end to implement this and multiple collisions with with less buttons from speeding
00:12:12
up the state of the art method at that time by
00:12:15
one other when it in fact i'd points five something which was a quite quite nice
00:12:21
uh there's always this isn't exactly thought because it's just really dying of computing exactly
00:12:26
simply things to a nice mathematical reasons but then say so not um
00:12:30
type of methods ah which are so exhibit twenty to to something where so you're so so the
00:12:35
point is to just to press if you with the content of this box if you tease
00:12:40
different part of uh so it's just a for a eye surgery on this is a search you can uh instrument
00:12:47
on the what what you try to decide is the the boxes rickety that's a certain
00:12:52
part of the of the instrument i'm going to detect that arises out for classes
00:12:55
on the what happens is that you you when you info on system which is supposed to look at the
00:13:00
image and then at some point make major conclusions okay this is like the the cheapest uh the instrument
00:13:06
but what you can do is that you can see that very quickly you have looked at enough evidence is
00:13:11
to reach a conclusion you don't need to one the competition companies competition looks like uh some of a
00:13:16
lot of schools a lot of hope at all the channels and again and again and again
00:13:20
well it's uh it's in the concept is quite easy as soon as we see that was cost that stopped if file from
00:13:26
the time we stop the competition and say okay i don't need to do it could be i knew that i will
00:13:31
it it's a bit like when you when you have an election on you don't need to also deserves to know that you know if you want some countries
00:13:37
and and eighty ninety seven votes for the employees president we don't need to see
00:13:42
all the votes to to to know that uh the game is over
00:13:45
on the the point now to implement this in a in a proper manner statistic alien species down but that
00:13:51
didn't get issues but this this worked uh mm dutifully again cutting the competition by opposed to another to
00:13:59
um now what we're trying to do is to controls i mean we could plant
00:14:03
because it's in the issue in a deep so courtney brown is a shipper
00:14:08
really angry so g. p. use knows who's fancy the
00:14:12
devices we have four apart competition have tens of
00:14:16
gigabytes of memory but it's actually a funny meeting fact also when we we had the uh
00:14:21
line of work well we tried to articles deep learning things out be metrics products so
00:14:28
you have huge mats we see that one thousand by one thousand or even more
00:14:31
on what you can do is to on could just matisse these metrics as a product lines but special
00:14:38
kind of product but that's the end of the day it's a product of small smattering sees
00:14:41
well that's what we do so instead of it being full metrics we have so called clinical product that might lazy so
00:14:47
that a number of parameters is fast more um it's actually closer guys some of the the sizes you know
00:14:53
on the show that's not a news remote who's been discount for but it has a
00:14:57
nice uh statistical properties controls are so good capacity of the model of anywhere
00:15:01
only show that and can you tell benchmarks for that particular kind of network would like your non networks
00:15:07
rubber to each state of the art was a fraction of the parameters to nominate harder so
00:15:15
does clinical product can be implemented was it takes you know to the
00:15:18
fast weight on something else would beneficent to be a computed
00:15:23
on the nose no an application where we equal things i mean we could twenty ski the i have this
00:15:28
uh project with a space on tonight to p. f. and collaboration with the university of berlin so
00:15:33
the goal is to design a new algorithm to construct a showcase of mass on the
00:15:37
images taken by yeah and the muscle data on so now such the shining methods for those those those
00:15:44
objections are still construction ah again uh you know uh networks on say surveys impressing which is
00:15:53
uh there's a simple things to to lose the bigger the major can push so you network
00:15:58
the better the desert because when you do so so a construction you get what nifty matron
00:16:02
right image and you have to reconstruct that's not so the distance of the point
00:16:07
on the videos the major battles the you can do a job because uh you can solve a
00:16:12
on duty so for instance if you're very flat tire out it's it is often skate
00:16:16
you need to see a very extensive contacts to to to match the point is always on duty
00:16:22
will know exactly the distance if you can't uh find details is uh is uh the major
00:16:28
that comment was images on heels the the digits of the students so
00:16:34
very skewed in designing our networks was to to sit in some way to count to the auditors onto
00:16:39
the moves and used as ways of information so that we could lead using we could paint
00:16:44
one in san just doing just doing these we were able to use a new if within by close to again one order of magnitude which
00:16:51
simply allows us to push during it into network on to beat the existing data
00:16:55
so this is going to under the viewing uh_huh nice conference on we are
00:16:59
quite happy with it because the we h. we beat state of the art and it's a very active domain
00:17:07
a note to tools of programs one key problem is uh which is uh we supervision to avoid the need
00:17:14
for detailed leaders so when we interact with the about the training
00:17:18
goodness uh when we when we interact with the undiscovered
00:17:22
partners there very often say they don't have they
00:17:25
thought they don't have labelled data so being able to get a small data sets its key
00:17:30
on here what we did was to train the so called matrix so to be able
00:17:34
to to compare to a person images without having a lot of images of each
00:17:39
simply by training in the comparison matrix so we have to to brought to images cause what does and doesn't
00:17:46
for the mouse uh oh construction we we were able to train the system without putting on clothes
00:17:52
by using constraint on the on the topology of this uh not still face so you know that spoken
00:17:58
with all these you know that's the flat i as i usually have this density of this
00:18:05
division so you are able to train we get liza network without savings and
00:18:10
one comes on there so we currently are looking at ways of
00:18:14
using a p. train networks whichever it does network which has been trained on it a lot that i said
00:18:19
are you able to use this information to train in those ah network and those up
00:18:22
past the smaller that that's a one here we are deeper new methods uh
00:18:28
where we instead of just outlining the responses of the networks we aligns yeah but it just opens i suppose
00:18:37
okay i would pass these so this is a long lasting a collaboration with e. p. f. and where we are
00:18:42
small choice that stanley nice to show that we're also interested in these yeah it's not that or
00:18:47
even it's a simply we we have an explicit model teenage even with nothing to put it
00:18:52
on then we either to use mean feat moderating might mean for documentation to
00:18:56
we compute the white europeans of persons given the images without much unknown
00:19:02
okay other meet about a a real project so this was the first undiscovered project that we
00:19:08
had reason has been affected yeah putting them on which one and both by an ass
00:19:12
so they came to us with the problem of doing facing notification we on phones
00:19:17
on the key element in face notifications to do very accolade face
00:19:21
localisation days of very nice technique for these so called uh
00:19:25
for i stuff a vacation trees um but supplements this methods the more
00:19:31
that was one hundred megabytes on as you may know if you
00:19:33
have a smart phone you are not very happy or you don't know the map which is more than like five or ten megabyte
00:19:39
so they came to us and so okay how can we solve this one interesting you were able to to
00:19:44
texas more there's in which are a series of trees with a lot of a big told us alley
00:19:49
space out big suspect although any doubt whether to use um i'd emotion in dimensionality reduction method that we had
00:19:57
evoke a completely different application exact case on the uh on it using against um but that that one order of
00:20:03
magnitude so in the end it wasn't i get because we were able to complex to less than ten megabytes
00:20:10
this is a project through the eyes are also uh the company uh in valet uh with the most earnest get has now
00:20:16
when they came to us because they wanted to find a way
00:20:19
of aligning or caps uh in manufacturing process is so
00:20:23
uh they say i've been making plastic tubes for four years stop reading actually machines that too
00:20:31
when they came to us because for some make marketing reasons customers wanted
00:20:35
now to have a caps which are not screwed but it
00:20:39
on which i did a constraint in the in the process which is you have to align the cap uh you know get amen oh
00:20:45
let's go to difficult methods with a good if your problem for many reasons but at the end of the day
00:20:50
um what did you not get out to the to the formation of the cat you
00:20:53
to the keeper itself was a major issues so you would not have a simple
00:20:58
kate i and because the cup with the from what we develop as a method where you put a in instead of every time
00:21:04
you change the cap model on the you have to have a technician going on the side stopping production for four hours
00:21:10
setting up a laser that is going to be a got twins the argument is quite
00:21:15
what we did was to develop a method where they put a bunch of caps like twenty of them and say now uh in the machine
00:21:21
they just sit there and the system the object apps compute the others a profile of
00:21:25
the caps that taking into account was it the formation on the provided separately gap
00:21:30
one in zen just think he's putting drops placing there and then in that or two
00:21:34
minutes that were able to calibrate and machine for for new model of debt
00:21:38
well there was a crappy so this is deployed this is so rude to the customers so the idea is
00:21:45
and now we have an ongoing project and know that it's finishing
00:21:48
on detecting it's to to defect uh automatically so you you
00:21:52
you wield you well some of plastic and you want to know if you if you have a hole in your process
00:21:58
and finally the last project which is a on going sings um so this is a company from the young
00:22:05
don't go and they are the readings habits which have solar panels
00:22:09
on the two with that um um um you do
00:22:12
on the what they do that's a localised to locate a bad weed
00:22:15
in the field on bass player beside the which was really cool
00:22:19
because you would use amount of heavy side by like yeah eighty percent to buy that you need computer vision to do this properly
00:22:26
on the one we have had so far to amass justices with them um
00:22:31
first one to compensate the quality of the image because you have a strong you mask on violation of
00:22:37
uh illumination and the shadows on the increasing between the reading
00:22:41
of the training set on spectral images you do is
00:22:45
i hold a whole do so we also we have a county uh we
00:22:49
are improving so the user to detection itself so it's easily name age
00:22:53
uh this is a production of uh what is the what is uh well
00:22:57
also collect we don't about weeding right on this isn't one comes
00:23:00
so this is our very very nice people okay i'm any less line i'm i was supposed to be
00:23:07
to be twenty minutes i'm doing it also just to to our conclude so
00:23:11
uh the the themes of future theme c. main uh improving
00:23:15
the practical use of deep running loose ski axes so
00:23:20
sings a need for data are controlling the competition uncle things and then we could plant
00:23:24
so we it's it's really on this is exactly t. shoes bows so italian impacted
00:23:30
also something which is showing up at the hideout quite heavy on that we have a huge project
00:23:34
uh in in in a in preparation for this is a notion of class an addiction intelligence
00:23:40
which are very i touches maniacs plus pixel fairness
00:23:44
ah he's a way i'm being nice is
00:23:47
following ethical principles um can you understand what is doing if u. t. pro and everything
00:23:53
goes well is it going to be associated with the program on the also
00:23:57
how can you be sure ones are long term that they eyes going to to do what you wanted to do
00:24:02
so we still specification of that's what i should do actually quite in
Conference Program
Introduction by Hervé Bourlard
BOURLARD, Hervé, Idiap Director, EPFL Full Professor
Aug. 29, 2018 · 9:03 a.m.
916 views
Presentation of the «Speech & Audio Processing» research group
MAGIMAI DOSS, Mathew, Idiap Senior Researcher
Aug. 29, 2018 · 9:22 a.m.
16953 views
Presentation of the «Robot Learning & Interaction» research group
CALINON, Sylvain, Idiap Senior Researcher
Aug. 29, 2018 · 9:43 a.m.
9721 views
Presentation of the «Machine Learning» research group
FLEURET, François, Idiap Senior Researcher, EPFL Maître d'enseignement et de recherche
Aug. 29, 2018 · 10:04 a.m.
14112 views
Presentation of the «Uncertainty Quantification and Optimal Design» research group
GINSBOURGER, David, Idiap Senior Researcher, Bern Titular Professor
Aug. 29, 2018 · 11:05 a.m.
3210 views
Presentation of the «Perception and Activity Understanding» research group
ODOBEZ, Jean-Marc, Idiap Senior Researcher, EPFL Maître d'enseignement et de recherche
Aug. 29, 2018 · 11:24 a.m.
5620 views
Presentation of the «Computational Bioimaging» research group
LIEBLING, Michael, Idiap Senior Researcher, UC Santa Barbara Adjunct Professor
Aug. 29, 2018 · 11:45 a.m.
4130 views
Presentation of the «Natural Language Understanding» research group
HENDERSON, James, Idiap Senior Researcher
Aug. 29, 2018 · 2:03 p.m.
8979 views
Presentation of the «Biometrics Security and Privacy» research group
MARCEL, Sébastien, Idiap Senior Researcher
Aug. 29, 2018 · 2:19 p.m.
6512 views
Presentation of the «Biosignal Processing» research group
RABELLO DOS ANJOS, André, Idiap Researcher
Aug. 29, 2018 · 2:43 p.m.
4029 views
Presentation of the «Social Computing» research group
GATICA-PEREZ, Daniel, Idiap Senior Researcher, EPFL Adjunct Professor
Aug. 29, 2018 · 2:59 p.m.
7159 views
Artificial Intelligence for C2ISR (Command and Control, Intelligence, Surveillance and Reconnaissance)
Olivier Volto, Airbus
Aug. 29, 2018 · 4:05 p.m.
322 views
Recommended talks
Presentation of the «Machine Learning» research group
Sept. 1, 2016 · 9:35 a.m.
9363 views
Presentation of the «Natural Language Processing» research group
Sept. 1, 2016 · 2:25 p.m.
1989 views