Player is loading...
Reality, Truth et Artificial Intelligence
Prof. François Fleuret, Professeur Ordinaire, Directeur du groupe de recherche en apprentissage artificiel, Université de Genève. Professeur Titulaire, École Polytechnique Fédérale de Lausanne. Chercheur externe, Institut de Recherche Idiap. Cofondateur Neural Concept SA
Wednesday, June 15, 2022 · 4:30 p.m. · 14m 24s
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:02
still before sinking exactly about what would be the content on the
00:00:04
near to make to talk with a bit stressed because they think
00:00:07
that the the title was very omissions that finally i manage to
00:00:13
two i think the chip on two big issues with i as as the so
00:00:18
maybe first to to summarise very quickly which is more than the high so if you go back to what was
00:00:24
just sounds on the way of when you computer attention and we've come from the computer
00:00:29
to do this you have to devise right to so called one so you have to write a
00:00:33
v. detail the stuff uh operations to make it on basic it is a computer would follow this is
00:00:41
carefully so we're we i don't expect you know tighten good cheer this little example officials who
00:00:47
do just faint so the the visa divide those of the fun to go and take a number
00:00:53
just you you there's a computer okay you have to count on zero two and
00:00:56
press one two and and then you have to shake to divide on them thing
00:01:00
oh well so far a ways of doing yeah i was for doing this
00:01:04
this way of doing scenes so people were kind of time to feature what was
00:01:10
managed care way of solving problems on then writing full on some that way with very detailed operations of how
00:01:17
to solve that one so example would be just playing so we can in some wayne's new we naive way
00:01:24
thing that we do have to be so so we just try to understand to
00:01:27
see the future of assumes you if it pleases me and just simply that data
00:01:32
and then again this type of fetish antigens was very forman systematically see like to be
00:01:39
just divisive leasing a little pieces in intensive manner on the computer just loosely hippies
00:01:46
well they're ashamed agency lies in the fact that in many situations
00:01:51
you may know that there is a v. c. p. that would work
00:01:54
but just the c. p. t.'s comment finally wrote a lot and off on it so that we do
00:02:00
so the they seem for example would be imagine that i want to write a piece of software
00:02:03
that critics assisted brought pressure from the age of the question i'm not a cardiologist and that uh
00:02:10
no sin
00:02:11
give me but that assets alleviates the turning points again that's on the on the right
00:02:16
a huge at each point was one to passion on what we can see even if we don't
00:02:21
win this in get carter g. we see that's a some kind of evolution in the agent about pleasure
00:02:26
so they can always do stuff i said sunshine onto so case there must be
00:02:31
but i can find alignment whose whose main points on just try to piece
00:02:35
of software that does is but the lines undefined so my piece of software now
00:02:39
as to monitor switches the uses the that's got a position of the lines us
00:02:44
because of the nine hundred should can just play with those two
00:02:47
parameters until my line was good looks good meaning i have that ah
00:02:53
on my line worsens data i still don't know any scene gathered g. i. signal i
00:02:59
mean scene the scene but now i can tell you okay if you're sixty years or your
00:03:03
take the pressure should be around this value but of course receiving additional support because not only is there is noise
00:03:11
probably hazel eyes actually on the nice in the measurement on some aspect of the
00:03:16
blood pressure can be credited but probably once in a lot of information on the status
00:03:22
jean jean because of the passion on we so we see there is
00:03:25
a constant but don't even something cases with gazes point of the yeah
00:03:29
we see that art is pretty outside and but still it's interesting that
00:03:33
even though i don't know scene i don't yet know any of the causal
00:03:37
process that's the late th and on the oppression i can already do some kind of tradition
00:03:43
on this kind of approach well devise a piece of code words out of sleep on it or something feeds
00:03:50
upon it doesn't that make it work kind of thing called a little bit what biology that so you have
00:03:56
blueprint of governmental system continue the any on then your
00:04:00
life and interactions environment will fix some p. parameters that
00:04:03
to be a human to so this as much of it is the development of enough where there that follows is
00:04:09
the person who now it's it's a whole lot on networks we have a piece of code is a
00:04:15
lot of the parameter that would fit in that out on this really what seems to be we turn on
00:04:20
works actually very well so you can difference and devise a program will
00:04:24
undertake doesn't put an image i don't put the classes object in the image
00:04:29
or even better you want to on them value lonely produce now that's exactly the same
00:04:35
so this leads to count models which are absolutely gigantic so it it went on
00:04:41
few tens of thousands upon us to know the last month from
00:04:44
today was one point six trillion parameters so it's completely crazy uh on
00:04:48
our also into says since those two thousand eighteen we're starting to see
00:04:51
who's withers which are extremely efficient for natural language processing would come back
00:04:56
so this of course is variable teleport captivating not not so good monitor exemption on networks with that of people on it does
00:05:03
is very efficient in many different application fee so you just a few station so it goes on
00:05:07
joe interesting missions were you given your few images from the camera architect to please see the shape of the scene
00:05:13
seen understanding you've given you any major for complexity scene and i played it
00:05:18
or the objects present and to see what what would say yeah image captioning unit and they would use
00:05:24
it as needed to take that risk of the content question answering so we get to large text on
00:05:30
information and then the system can answer questions on figuring out which you symbols because she the text
00:05:37
good planning so means that you games on even now complex controls exactly some work where
00:05:43
the fusion new graphic and she's basin confining up last night's x.
00:05:48
m. unstable computed compulsive warren has been advised to efficiently with this method
00:05:55
so the the first point so this is what was
00:05:58
okay well that there are two fundamental for the future you should
00:06:01
be set to the first one that yeah that's not understand reality
00:06:06
in the sense of the the first part of that my example is proportional was nice
00:06:11
the line wings that that will happen if i say so
00:06:14
i expect to be moment it really complicated stuff so we'll premier
00:06:19
aggression so now when a member might cools you have printed as it can be some of the other side or the other
00:06:25
uh so yeah the line would be put in your of the d. v. one but i could goody two three four five
00:06:32
but it's up to me a a pretty near the nine would
00:06:35
go even but also the green dots but we see texas and so
00:06:39
we now new with your brain you look at this with the i nine right now that makes no sense you can the big question from
00:06:46
but from the perspective of the train it's a really good food
00:06:49
because they were supposed to be point so we can uh proves oops
00:06:57
we can have who've mathematically a very nice result which is that if you have a set of input
00:07:02
is on the you look at your data would be quite a that that seems to be fitting data
00:07:08
if you we see but he's you also have a long ways
00:07:11
describing the put disease we can we mathematical down which state that
00:07:16
the shop as a description of the creation choose the better the notation to
00:07:20
that sixty she's so classical fussing we need that they expected something something we
00:07:26
she wanted you shouldn't mean to guidance design so if you have two expansion of same
00:07:30
power the more concise once is a bit uh so discourse status regional networks in aiding them
00:07:36
on quintuple brought a addicted by yes you watch over there
00:07:41
to not shiny bees or problem oh there for that when you have a ten
00:07:47
also in fact is whose vision network so as you can on
00:07:51
networks days uh would we don't or something l. e. y. but
00:07:57
somebody's we see that they are apparently bias for that kind of time it's so
00:08:02
it from this perspective you started it'll be it's a weekly light ram disk any foe
00:08:07
but even in that case even if you have a perfect and divide
00:08:10
it so your mobility model anywhere that's not so process the way the models
00:08:16
the way you but it's been a minutes but you have to get in and has nothing to do with what would be
00:08:22
your physical explanation so we we created a recluse grease monkey f.
00:08:25
a. we created a company a few years back whose business he's
00:08:30
we we try to find more there is to meet expenses innovations so normally
00:08:35
if you want for instance computer which of you are we in for great
00:08:39
you you construct is thinking that if you win on the new one haiti competition which are
00:08:44
by you that immense value applied when the rage of the physics of three on it on to you
00:08:49
have to have a competition you get an estimate especially what we can do is to train a neural network
00:08:55
to get the shape of the windows input on directly paid you
00:08:58
would use special interest anyway so it supplies and surviving which was anyway
00:09:02
but that seems a competition which a network as intimately as innocent
00:09:05
to do what's whereabouts physics so it i mean it very well but
00:09:11
physics knowledge say it's sue the on product is very the special food is very red meat
00:09:19
but it does not capture at or what you would capture as if this is
00:09:26
also now resigned that understanding i didn't t. know networks also
00:09:29
start to mess it'd be tweezers or understanding of truth so
00:09:35
the first thing is that we can train used to use them with their
00:09:39
stupid use you that that so we have the pitch of strain on the left
00:09:42
ah pretty generated so typically you have shown to non networked everything set
00:09:47
with hundreds of thousands of features of the analysis of the each other nations
00:09:52
on then continue over there so that the g. i. usage features of
00:09:55
course features except uh you know i think that the buttons or something
00:10:00
both the floppy chosen the left on the two faces and the button on the
00:10:03
top right are completely fate so sci major stress on the subject we're going to see
00:10:08
we'll just generated by a more there they tried to meet the statistic
00:10:12
of images of faces and you got versions of your images of of past
00:10:16
on on the bottom right there it's so good effect where you use a neural network
00:10:21
to replace the face of an actor which was the which ended with the ones left
00:10:26
by the face funnels august which is produced and the right one says don't exist on
00:10:30
on our network so this actually was not done by uh the fancy indigent agency form so
00:10:36
the ah country but uh it's actually done by somebody i would seize
00:10:40
basements we don't discuss and under the distant on network and played around
00:10:46
on the side missing uh missing a little bit with what is truth
00:10:50
the truth that some of the exhibit is often problematic on it since
00:10:54
and we all know what happens when yeah so yeah what they show i support of so called language models which will be great at the moment
00:11:01
on those two small that's work as follows so they are trained given a piece of text to play with the most likely next work
00:11:07
so what you can do is to start a sentence on that wasn't that will do the next world and that's the next world and the next one
00:11:13
on next and the next one on the summer that's all trained and g. don't call because of that that which ah any for me to be the ah
00:11:21
for once the discussion bought and stuff so they say
00:11:25
is it true that they capture is simply what people say
00:11:28
well we knew that what we're seconds it's like the the segment result from so we are just as an instruction
00:11:35
the board of directors also once i tacked on the following about the part that was produced by the mother
00:11:41
well as you can see so since was that that training set ah probably
00:11:46
from the fifties we see some by a string of that may not be the
00:11:50
yeah but in in the deployment of the systems of friends and so that's why
00:11:54
comments guy is to continue to them will continue these are that's in on skis on
00:11:58
the best foreman sky is to be a housewife so we we probably don't want
00:12:03
to us kind of rules actually be light on to the pros assistance in the injury
00:12:11
another example is kind of true so some other cyber to upstate so easily
00:12:17
low resolution one can read it gently goes like like in in t. v. input is t. v. series
00:12:25
that you can have scales alliteration so if we stuff on this image that i
00:12:29
think we all abstaining you know with your brain this is what the model that's okay
00:12:33
so this this nation's right if you don't see you back the image on the left
00:12:37
but of course we all know that on the left it was of them are but the
00:12:40
system was trained on so many and that doesn't so much bias the while white young people
00:12:46
that's the most likely on storefronts model when you say
00:12:49
the loader teenage he's away young guy that we seem to
00:12:55
finally i don't know if we see these but we start given to have a point of truce as
00:13:00
an escalating as a human so no language model but that's not really able to pass the turing test
00:13:06
maybe not completely the topic was bassett so yeah maybe we don't have time for this but
00:13:11
uh you you can go around which more they're always to explain jokes on actually
00:13:18
maybe maybe maybe you have to download the side adjusting the scene
00:13:22
it's actually view some some other just by editing the next local body
00:13:26
he simply able to give use expansion of joe which is something
00:13:30
that seems really human on finally regarding the confusion between a human mine
00:13:37
on the i said stan wouldn't to some others which are
00:13:41
train only used at a set of images and text so
00:13:45
you really asked me and what me and sites actually is the number of pairs of image and
00:13:50
text on than what you can do is to train with us that if you give the text
00:13:53
it can produce an image so two stages local used by the by an option on networks
00:13:58
so to first walk would use if you didn't want the painting of the last they see the teaching right
00:14:03
on the two at the bottom up would use if you ask it to produce images of some
00:14:07
other so those images are converted generated by you know on it you meant in the planes see
00:14:14
okay that was my age
Conference Program
Welcome words
Aurélie Rosemberg, Prof. Antoine Geissbuhler, Vice-recteur de l’Université de Genève
June 15, 2022 · 4 p.m.
Conférence opening
Mme Fabienne Fischer, Conseillère d’Etat du canton de Genève – Département de l’Economie et de l’Emploi (DEE)
June 15, 2022 · 4:05 p.m.
Evolution of Computing and Vision
Prof. Giovanna Di Marzo Serugendo, Directrice du Centre Universitaire Informatique de l’Université de Genève et Professeure ordinaire à l’Université de Genève
June 15, 2022 · 4:15 p.m.
Reality, Truth et Artificial Intelligence
Prof. François Fleuret, Professeur Ordinaire, Directeur du groupe de recherche en apprentissage artificiel, Université de Genève. Professeur Titulaire, École Polytechnique Fédérale de Lausanne. Chercheur externe, Institut de Recherche Idiap. Cofondateur Neural Concept SA
June 15, 2022 · 4:30 p.m.
117 views
Governing the rise of Artificial Intelligence
M. Nicolas Miailhe, Fondateur et Président, The Future Society (TFS)
June 15, 2022 · 4:50 p.m.
Artificial intelligence & issues related to States, Peace and International Cooperation
Dr Giacomo Persi Paoli, Directeur du programme de sécurité et de la technologie de l'Institut des Nations unies pour la recherche sur le désarmement UNIDIR
June 15, 2022 · 5:10 p.m.
Panel discussion
June 15, 2022 · 5:30 p.m.
Closing of the 50th birthday celebration
Jean-Pierre Rausis, Président de la Fondation Dalle Molle
June 15, 2022 · 6:30 p.m.
Recommended talks
Réalité, Vérité et Intelligence Artificielle
Prof. François Fleuret, Professeur Ordinaire, Directeur du groupe de recherche en apprentissage artificiel, Université de Genève. Professeur Titulaire, École Polytechnique Fédérale de Lausanne. Chercheur externe, Institut de Recherche Idiap. Cofondateur Neural Concept SA
June 15, 2022 · 4:32 p.m.
143 views
Deep Supervised Learning of Representations
Yoshua Bengio, University of Montreal, Canada
July 4, 2016 · 2:01 p.m.
2372 views