Player is loading...

Embed

Embed code

Transcriptions

Note: this content has been automatically generated.
00:00:00
i think survivor introduction and thank you everyone for being here
00:00:05
i'm gonna be spending a few minutes talking about social computing and illustrating my view
00:00:10
of social competing with concrete projects that we develop over the last two years
00:00:14
but before that i would like just to start describing what it is so so you're comparing us a disciplined really integrates
00:00:21
oh no mention morals from a ubiquitous computing machine learning social sciences with two large goals
00:00:27
the first one is about insights know the words to understand how people
00:00:32
use participate in ultimately benefit from social technical systems
00:00:37
like social media phase balking sacraments own
00:00:41
well they're online systems for instance cover a crow oh sourcing
00:00:46
online companies but also more recently also my systems that link online
00:00:50
activity to the physical world like oh we're already in b.
00:00:54
that's the first go and then the second one is based on the understanding and those inside
00:00:58
the can be derived from from analysing the systems how can we build computational methods
00:01:04
that can all made certain tasks tours a better understanding of these phenomena
00:01:10
but also trying to automate tasks that could support what people do on those online systems
00:01:15
so in all these do a keyword use engaging engaging people with technology
00:01:21
as we all know we can do a lot with a online systems because we live in the network
00:01:28
and i think the fact that the our physical life in our online life are more more intertwined
00:01:34
make it very clear that we can essentially do most every activity
00:01:37
you know did you live through also should incur systems and
00:01:41
the social meat is clearly the quintessential example of of a system like this
00:01:47
they're kind of activities with the one face broken inside grammar trader how actually shaped
00:01:52
recently for better or for worse some of the last uh trance inertial side
00:01:57
so much of the research done in social me has to do with looking at uh
00:02:01
oh no so you're comparing has to do with looking at social media as a
00:02:05
source of data but more importantly as a as a source of of our research questions
00:02:10
but as we have also seen in the last few years there should the limitations of
00:02:13
thinking of social media as a uh so fair description of society for several reasons
00:02:20
i refer to them as biases but there were other ways of all conceptualising the limitations
00:02:24
the first not only is this is ergonomic one which is not not everybody
00:02:28
user percent did not remember societies are presented for your social media
00:02:32
the second one is that as a result of that not everything is represented as social media
00:02:37
and if it relates only to the geographic aspect of it but the replays
00:02:41
is represented insertion media and has been a a lot of uh
00:02:45
press in the last couple of years about the facts of all these blind size that we have
00:02:50
when we take these as a behavioural traits of our society and we
00:02:53
try to understand our society from these very biased source of information
00:02:58
so another way of thinking of how to contribute to if you want to cover some of the blind
00:03:02
side of what social media can give us today to try to understand societies online is sucrose source
00:03:08
or sourcing in which we're gonna engage specific groups of people to donate data it so that they don't try
00:03:13
to either look at these my insides and get some clarifications of some of the limitations that exist today
00:03:18
but also to try to help research namely in the social sciences were in the behavioural sciences that's
00:03:24
for a long time have had the desire to understand very intricate phenomena everyday everyday life phenomena
00:03:31
but that's so far have been relying largely on surveys and indirect methods of observation
00:03:36
what technology brings us a great a possibility today is that because many of us are caring devices with us
00:03:42
we can actually be closer to that very fine grain description of behaviour the
00:03:46
social scientists have hard for a long time the wish to have
00:03:50
and combining parts of automation parsonage technology and the very sound methodology that exist in social science
00:03:56
we can uh we can ideally contribute to the generation remote not much
00:04:01
so in all these countries i'm gonna give you a three examples of research that i've done what we've uh collaborators in the past
00:04:07
our group right now is about a four to five people but typically can between five and ten people including
00:04:13
pages tunes balls out in collaborations and as you will see our research is very
00:04:18
interdisciplinary so much of what we do we do with people from other domains
00:04:22
i wanna talk about a project that is uh has ambition of
00:04:25
understanding nightlife through a crowd contributions called and asked to don't
00:04:30
the second one is a pretty good we call bytes and bits and looks out
00:04:34
eating behaviour and how we understand eating routines in the light for more data
00:04:39
infinite i'm gonna talk about this big plus from which is a social innovation uh attempt to try to
00:04:44
uh make all the people who are not necessary technologist potentially beneficiaries over the work we do
00:04:50
so let's start first with the dust don't project
00:04:54
what we're aiming to do a very concretely is true through our multi disciplinary partnership that includes
00:05:00
a social jog refers from the university of zurich and it be limited use of alcohol consumption probably shelf
00:05:06
where before in addictions wheezing that are now at the laptop university in australia
00:05:11
our goal is to find a framework which we can capture
00:05:15
in understand urban nightlife patterns of us uh suisse youth
00:05:19
the possibilities for this is such a sensor very large because you can imagine that use i exposed to certain number of risks
00:05:25
that perhaps the privileges of consumption of substance it would like to get more insights about
00:05:30
job reference also care about the right this is used in how they use the spacing how the appropriate the urban space
00:05:37
and for us computer scientists were interested in uh looking at the data mining aspects
00:05:41
of it and how we can also build models to support these uh research
00:05:46
so what is important about this project is that there's always a part has to do
00:05:49
with insights in understanding in a second one has to do more with automation
00:05:54
for this particular project we uh involve population of over two hundred
00:05:58
young people between a sixteen in twenty five years old
00:06:01
from the two major nightly hops in switzerland namely saw rick anderson and these population was
00:06:07
um voluntarily willing to share data with our research team so our phone application
00:06:12
uh in the was very specific uh with respect to its use we require them
00:06:17
to volunteer information on friday and saturday nights between eight human for yeah
00:06:21
the idea was to get a snapshot of how life uh at night he's experience by young people in these two six
00:06:28
more specifically the framework that we have built which is uh not is a combination
00:06:33
of sensing so the the different sensors and you can record on the phone
00:06:36
location accelerometer orientation battery and so on
00:06:41
combined with sort of a is that a participants in this study would have to feel themselves that would
00:06:47
include for instance telling us where they where telling of the social context in which they wear
00:06:52
but also is more specific for research in alcohol consumption what kind of brings the where uh
00:06:58
uh drinking doing that particular moment in addition to that we're going to take photos of the drinks they took
00:07:05
and as in whenever he was appropriate to take snapshots of the of the places in which they were
00:07:10
so the idea here is really to combine both so shall uh
00:07:14
interaction in other words are relying on people's natural uh
00:07:19
brad these like they would want histogram of taking photos and videos of how they
00:07:22
go about the the there live during the night dinner but also having
00:07:27
in the bargain a phone that would sense sensors yeah since we're data that
00:07:30
would let us uh uh essentially correlated to the kinds of information
00:07:35
no i'm gonna give you a very a quick overview of some of the research questions that some of the part as having
00:07:40
interested in and you know answering in some of the findings that typically we can find with these kind of framework
00:07:47
very basic one is this question on uh where do people go right and when you think about social media in use what you tend
00:07:54
to think about bars and restaurants and and nice places where where
00:07:58
life can be active routing there's this entire discourse about how
00:08:02
normative or social media is in hopper formative also socially it can be
00:08:07
so it turns out that vector analysts people who are not really performing assertion yeah but i just giving data
00:08:13
as they go about their their nightlife which are so that uh will have all the places in which of these sixteen to
00:08:18
twenty five a young people go our private place you their their home or some of the parents or formal difference
00:08:26
and this is quite interesting because this is not the double information that you would normally find all social media sources
00:08:31
not another interesting aspect is that uh from dollar fifty percent is
00:08:35
a public space roughly half of that are non commercial places
00:08:39
which in these concrete case would be barks this street certain squares the like the the lake sites and so on
00:08:47
this is interesting for for uh the european context because interested in particular
00:08:51
really don't just read is not criminalist so people actually can drink like a whole it to spend time outside
00:08:57
and uh but it does is very interesting about such a the the economics of being outside didn't like that
00:09:05
and that means that not everybody has enough resources to go to a club every weekend every time because i've
00:09:11
these are really can be a shown we
00:09:15
uh with some statistics by the main important uh finding for our contacts in exercises was this
00:09:22
um confirmation that young people really spend a considerable time aware for
00:09:26
from commercial mightily areas with implications of that can have
00:09:30
for civil disobedience and also there were substantial differences between what we found using
00:09:35
colson seeing compared to what you would find in a socially just
00:09:41
another example where we start to makes more the information with the social science part is this question
00:09:47
on can we actually capture basic things about places how bright are there or how loud right
00:09:55
if you wonder why is this important well i asked a
00:09:58
superset is my have some hypotheses about certain behavioural
00:10:02
um trains that might be mediated by certain environmental characteristics
00:10:07
for instance the people bring more because they are in a louder plays or
00:10:10
the people do certain things if they are more you know quite please
00:10:14
purely because we have the video we have access to images and we also have
00:10:18
access to the soundtrack so we can automatically extract basic measures of brightness
00:10:22
and uh of loudness and then categorise these concrete measurement with
00:10:27
respect to the different categories a place where people go
00:10:30
so just to give an example we focus on the diagram on the left you can
00:10:34
see for these measure of loudness they hire the the louder that uh without
00:10:39
much surprise clubs events in bars i'll just as a measure uh objectively so to speak by
00:10:46
by the machine and a pair of places are the quite as one but if you also look
00:10:50
at the specific a case of private place it was it was how much like the variance
00:10:56
and what does that a larger areas means is important
00:10:59
also for social size because essentially just signals
00:11:03
that a pretty place might have a multi function use another word you could
00:11:07
have a quiet night with your friends with your boyfriend or girlfriend
00:11:12
in a very simple environment but you can also use your probably home for part in which case the level of loudness could increase significant
00:11:18
right so these basic finding can support for the research about the use of the private space by young people
00:11:25
now if we look at a case of the on the right side of the slide and we focus again a couple places
00:11:30
again unsurprisingly bars and clubs are the darkest places as well as this public space that is not commercial
00:11:37
because of actually look at the videos people tend to be parks completely at dark
00:11:42
or in the by by the side of the lake completely of dark and this is actually
00:11:46
interesting to capture in the database actually a signal how people actually spend their money
00:11:51
right in a in a quite environment in the formal isolation where they
00:11:55
can actually be free to uh do whatever uh that they wanted
00:12:01
now there are also other examples of research we can do on these kind of framework that are really more computational
00:12:06
it basically want to move from the perspective of places from looking
00:12:09
at more physical attributes to more psychological attributes of place
00:12:14
so more concretely one question that one could ask is what makes a place be perceived
00:12:18
i've been of a certain time for instance as a form of place was a cosy place was the dangers but
00:12:24
so it turns out that we did learning right now we have the possibility of trying to understand within an image
00:12:31
what are the reuse of the image that is the are used by the machine to make a certain decision
00:12:36
so these two images i'm showing here have been both classified by uh didn't work as being a formal
00:12:43
and then we can use a basic design decision uh i agreed and within the c. n. n. a formulation
00:12:48
that allows us to actually show that the air the corresponding these guys to ah the tablecloths
00:12:55
and then when classes are what the machine has been using is using to make these decisions
00:13:01
obviously a one could go very deep into into the so many try to understand
00:13:06
also the specific features of pacific subclasses of features it's just such a good point i ever
00:13:11
met the psychology in design but certainly something that the cube explore for in that but
00:13:18
if only another example of automation is this issue of could we actually identify
00:13:23
based on the sensor data which is something that we can collect where someone had drink
00:13:27
an alcoholic drink last night so the basic a framework is a simple
00:13:32
conceptually we can take data for instance from the entire previous night
00:13:36
location um motion low twos use all caps and so
00:13:41
on with that train a model that good ultimately
00:13:46
uh be able to respond a question but i call was consume in the previous night
00:13:50
it turns out that with very simple features and relatively little data this system trained this way can
00:13:56
roughly guess three out of four times where there that night included the consumption of michael
00:14:02
and if you might wonder why could this be possible way clearly the place to juggle the company which you are uh
00:14:08
the the the fact that you are a more popular places or louder places and so
00:14:12
on is informative to some degree of the specific activity of being a call
00:14:18
so i think these are just examples to try to to send the measures that thinking of processing this way enables
00:14:25
both new insights in the possibility of creating new algorithms that relate to places and activities in this particular context
00:14:35
the second example is if you want the variation of the same topic but is related more to every day uh you
00:14:41
and as we know of flowed really the the food we eat is highly dependent on men context factors
00:14:49
where we have a specific as the specific moment where we're by ourselves or not whether we're
00:14:54
working or not but he's also in france by psychological factors how we feel about food
00:14:59
how we perceive would've been healthy or not whether we know that eating certain things are certain time will be healthier now
00:15:05
so we know i in other words that all these factors that having studied largely nutrition science
00:15:11
and for which there are theories and dials observation uh studies that can tell us how
00:15:15
people actually eat in in real life but also how big conceptual ice for
00:15:20
things a simple of these theoretically as what is some you know what's the snack so all these body of literature could
00:15:26
also be a very basic then we have the possibility of using a phone application to which people document that
00:15:32
uh their daily eating patterns in out there with the analyses like the one actually
00:15:37
scratch so we have for this in collaboration with a mess the research
00:15:41
we'll uh an application called bytes and bits that uh was decided to be around in
00:15:47
that kind of community experiment with e. p. f. l. uh as a student population
00:15:52
so the the entire a procedure was based around the development of these application
00:15:59
so which in yesterday's will be invited to download ending contribute data for number of days
00:16:04
in specifically what people were asked to do was to whenever they were having some me a meal or snack
00:16:10
they could take a picture of what they were getting the good document again shorts or they would
00:16:14
they would tell us their social context have you other factors that were important for us
00:16:20
this is kind of the the standard for the who asked people to take in the document uh in more detail
00:16:26
so again it's running age group of students were able to collect this experiment about a two hundred days of data
00:16:34
the data that they actually report all or most of their yeah it in your size next
00:16:39
as well as the context of these yeah eating locations where wayne who why so
00:16:47
in addition to that people working i think it's uh we also have information about the physical activity because there's some
00:16:52
at least hypothetically connection between the level of physical activity you have a uh these are you have four four
00:17:00
and again i'm gonna show you just a couple of examples one about the kind of insight that you can get in is gonna
00:17:04
be a superficial because of lack of time in another example of what would be the more from the the a. i. side
00:17:11
so it turns out that if we look at the aggregated patterns of these hundred
00:17:15
twenty e. p. for students with respect to where they eat meals and snacks
00:17:21
which is these applaud them showing here you can see that home and
00:17:24
school by far are the two most popular categories for eating
00:17:28
and that there is a difference between uh where you eat your meals and where you yours next
00:17:34
if you look up the battery uh of the uh corresponds to
00:17:37
home so home is the most popular category to either meals
00:17:42
with strong signals that students have the breakfast home have the dinner home
00:17:47
ah in in country to that snacking is most common at school
00:17:52
because in it doesn't make sense because this is essentially what they spend most of the time
00:17:57
uh whether they are you can see clearly it's not a a binary pattern but
00:18:02
uh there's a a certain uh consistency in terms of their bit uh trends
00:18:08
and i would look at another basic button if you look at uh the social context in which uh people in these particular publisher
00:18:14
need you can see at least so personally to me that many of these uh these meals and snacks i consume along
00:18:21
and is my also reflect the specific demographics of the population the
00:18:25
fact that many people might even studios are small homes
00:18:30
still housing where they have the on space and they by themselves but is nevertheless interesting to see that
00:18:34
whether we tend to think about getting as highly social uh event it's actually not so much
00:18:39
right because you can see a lot of uh of uh forty to forty five percent
00:18:43
of these events that people have made themselves not this can have some implications for a student
00:18:47
palpably feel in in this is something that could be also taken over by our partners
00:18:54
now we look at television part we can do with it kind of they after we do this analysis
00:18:58
uh just for purposes of illustration we essentially a
00:19:03
talk basic features that can be extracted from the counterpart basic are characteristic like the
00:19:08
time when when something was it in how much time had passed and the
00:19:12
last intake and they where people where and what it was they were doing
00:19:16
what they're watching t. v. what they doing nothing else were they working
00:19:19
and with these we again bill a standard uh classifier to try to classify as specifically the location a similar is not
00:19:27
on the various or phase seems like a trivial task because you would
00:19:31
say well we'll have to be between and seventy nine and uh
00:19:35
or eleven one and so on but it turns out that how you define your mail sense not this highly i do some credit
00:19:42
that was what i define for a meal maybe for us not inviting me a meal for me to they might not be a mail tomorrow so there
00:19:48
is actually many sources if you call them noise but realise human behaviour that
00:19:53
complicate these apparently simple task of the fish to do the minutes that
00:19:57
was is more interesting what was i thinking about going be on these as
00:20:00
in can we start train machines to predict most likely moments when
00:20:04
people are willing to or likely to eat a meal or snack okay we think of other applications that are more more with it
00:20:12
so these again illustrates the the the idea that again within
00:20:18
framework allows both both inside information was is it not
00:20:23
no like to use just the last a minute of this presentation to see how we
00:20:26
can push some of these ideas beyond appeared academic domain and try to uh
00:20:31
uh haven't measurement we we that stakeholders be on the on a university or the residual that we have built
00:20:38
a platform that we call civic that is essentially apart from that allows people to run so called experiments
00:20:44
by which they can design surveys of we can design experiment would allow you to collect the collect video so images
00:20:52
it didn't it to get technologies we can actually launch gap for you and people can connect that they're gay
00:20:58
in this context we partner with a singular sign um in that two thousand sixteen
00:21:04
because the city had these a relatively or do you need to understand what
00:21:08
was the state of street harassment in this c. d. over sign in
00:21:12
there was essentially no previous statistics and they have to come up with with
00:21:16
the way of actually measuring the state of this problem with it
00:21:19
so sort of some serious of of course the sign sessions with this and
00:21:23
with them using the the civic application away obliging a basic mobile up
00:21:27
but then would be run by specific uh city employees that are known as
00:21:31
in my correspondence of people what what is the the white not please
00:21:34
why social workers normally work on on the weekends to talk to people
00:21:39
about the different uh there's a risks involved in my life
00:21:43
they were the ones in charge of collecting information and after the study in in the
00:21:48
this that is by analysis of the data which are now that well that the
00:21:52
present the problems of the problem was much higher than anybody having my didn't end there was sort of these was that
00:21:58
to motivate for the policy by the civil sent to actively deal with the situation which until the study was
00:22:04
i've done was not solely known by that enable these process
00:22:08
of thinking about uh changing something in in d. c.
00:22:13
so i think you for your attention if i can just get with the words that summarise what what we do would would
00:22:18
like to keep doing it is this for our research people is always at the centre in technology needs to serve
00:22:25
ideally the public interest i think this example uh hopefully show you what's what's the implication
00:22:30
of that finally working across disciplines is really necessary for the work we do
00:22:35
and if you have ideas for collaborations you speak about the if

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

Multiple ways of building a recommender system with ElasticSearch
Andrii Vozniuk, React-EPFL
11 May 2017 · 1:04 p.m.
Mobile multimedia services based on user attention profiles
Francesca De Simone, EPFL
1 Sept. 2011 · 10:03 a.m.