Player is loading...

Embed

Embed code

Transcriptions

Note: this content has been automatically generated.
00:00:00
i've seen several speakers this morning talking about the importance of interpersonal skills
00:00:05
so interpersonal skills are important of different parts of our work live there certain
00:00:10
important to get a job there was important to advance with a job
00:00:15
but i think it's also more important and that is really an element of our daily life
00:00:19
and what is important for this ah audience is to show a little
00:00:24
bit whatever suisse research allows are doing today with respect to
00:00:28
people in the workplace from the perspective of computer science and psychology
00:00:34
the good news is that there is a lot of literature in
00:00:37
psychology in a specific even verbal communication we should sure
00:00:41
that has set a very strong signs on what kind of information
00:00:46
we convey with our voice with everybody will ever face
00:00:50
and especially in the workplace is important because as i said many of the jobs we do actually are also skills
00:00:57
when we work force for services in when we essentially interact with customers
00:01:03
it is known in uh this uh scientific literature that ah or never communication is actually an
00:01:10
alternative channel we worked we say so not only what we say how we say
00:01:15
also that these signals we come they are often unconscious because we
00:01:18
typically have a limited coming to resources when princess in a
00:01:22
job interview when we are thinking about what we're saying we perhaps uh pay less attention to how we're saying it
00:01:28
and most importantly nonverbal communication reveal something about our state of mind
00:01:33
our personal traits knows about the degree of relationship we have with people will interact with
00:01:40
we have developed a research over several years now in collaboration with the social psychology
00:01:46
develop a multi disciplinary approach to understand good first impressions and how
00:01:50
we can use computers to try to automate some of
00:01:53
these processes in the automatic how to support users conferences are looking for jobs want to improve their sub skills
00:02:00
specifically we have been working on it yeah with our colleagues from the university of lausanne
00:02:06
and from colleagues from cornell university decision degree this approach that involves are ubiquitous computing
00:02:12
artificial intelligence of machine learning but also social organisation as a coach
00:02:18
we have worked closely also with actors in the soft skills industry or if you want the the
00:02:25
hospitality industry specifically for the program we are we talking about
00:02:29
these findings by this was national science foundation we can work with the hotel school
00:02:35
but essentially stitching student how to project was first impressions developed was of skills
00:02:42
so from the perspective of what we're trying to do from the research um objectives we can name them
00:02:48
as the following first what kind of behavioural cues that relate to you know positive first impressions
00:02:55
can be extracted by automated means and we'll discuss this in a in a few minutes in more detail
00:03:01
second is really how can we use this information extraction of these fair cues that are about behaviour
00:03:07
to be able to train students with the train novices on
00:03:10
how to improve based on information they a show car
00:03:15
and we have studied this in multiple situations to try to appeal to the
00:03:18
real live conditions under which will work specifically focusing on job interviews
00:03:23
and on a service encounters more specifically on reception desks for a hotel employees
00:03:33
from there i would say most big picture perspective we're actually here with two main challenges the first one
00:03:38
is a human size channel i i check challenge in the second one is a computer science work
00:03:44
the first one is really is how do people perceive others so this is a subject on an
00:03:49
overall communications or psychology organisational psychology the second challenge
00:03:54
is how do we connect this to computing
00:03:57
and i'm making their offices here over and over again that you cannot achieve well what we're
00:04:02
trying to do without the two disciplines you can apply technology in avoid you will fail
00:04:07
or you could the celery but also we not being able to proceed to real life if you just
00:04:13
a lot so the the the lesson here is the multi disciplinary approach to do these guns
00:04:20
so i'm gonna have to walk you through a couple of of the main challenges on the human science park
00:04:25
and the first one really is how how do we perceive others for instance in a job interview
00:04:32
what kind of the inferences we make how long it takes to make the inference how accurate just those inferences are
00:04:39
psychology is know very well that ah in a situation also situation with my elysee
00:04:45
and number of constructs that we very rapidly make decisions about others
00:04:50
it and sometimes the decisions are gonna be correct in some in some cases they are not gonna be
00:04:54
that's correct differences in the case of our reception desk that you are customer coming to hotel
00:05:00
well you would probably make very quick assessments on friends of the professional skills of the person who is behind the desk
00:05:07
how multi very they they look at how competent they really feel to you with respect to the job itself
00:05:13
a second part would have to do with their communication skills but was how
00:05:16
how clear they are how persuasive they are when they interact with you
00:05:20
i thought i mentioned would be a general social skill well related
00:05:23
to how positive they are how enthusiastic they appear to you
00:05:28
and obviously if you were more in a situation where you have to assess the
00:05:32
tractor to make a decision you would like to know for instance how
00:05:36
well they perform the job overall how positive impression the person behind it this
00:05:41
uh uh sent to you and all this information happens really the snap
00:05:46
and we over the don't ask ourselves about all the things that we were actually varying them as we interact
00:05:55
we have developed a a a a problematic framework through which we actually capture
00:06:00
students interacting in a so in a similar situation about the receptionist situations
00:06:05
and then we have of servers looking at the videos that show
00:06:09
the interaction in great the participant in other words employees
00:06:13
and what is really critical i'll to make progress in science with respect to
00:06:17
human perception is really to quantify in one way that he's valley
00:06:22
what the perceptions are because as you can imagine some of these concepts are subjective
00:06:26
but they are still important when he's a necessary is actually to quantify how much variance there
00:06:31
is in each of these concepts in whether the variances acceptable to make for the decisions
00:06:38
so as the same descended issue some numbers of these are numbers from a basis study
00:06:43
that shows how much people agree on a scale between zero and one
00:06:48
with respect one number of applicants with respect the number of dimensions
00:06:52
so you can see that the numbers are somewhere between point sixteen point seven five
00:06:57
that means that they're the degree that several people have when they look at
00:07:02
the uh candidates is that's a acceptable but is not extremely high
00:07:07
you could see some variations across their measures for instance positive impression which is
00:07:11
roughly our overall sentiment tends to have highest agreement when people i
00:07:17
rape uh candidates with respect to these image and it will be all
00:07:20
others that will have more variance which means people didn't agree less
00:07:25
this is obviously a important for psychology in and of itself but it's even more important when we use this
00:07:31
information as labels to train machines and uh it's important to keep these for the rest of the talk
00:07:41
another important part is once we are able to extract behaviours and we'll get to that in a moment
00:07:45
we can also link those behaviours we extract with the labels are the rates we gave two candidates
00:07:51
so you can see in this case for the reception desk we have found in our studies that
00:07:56
a a reception is that are perceived as more positive than to have specific behavioural uh batteries
00:08:03
specifically they tend to be more active visually they move they also not more in other words they
00:08:08
show more signals of uh of being aware of the director in in showing some feedback
00:08:14
net and then at the same time to have a certain coordinated pattern of speech respect to fluency
00:08:20
and with respect to essentially keeping a good pace in the way they talk what's interesting though is that researchers to show us that
00:08:28
our make inferences or their candidate one also can look at the instructor in
00:08:33
this case the customer review was a job interview at the interviewer
00:08:37
and this is quite a quite important because it shows how behavioural yeah
00:08:41
in this case is the the result of interaction and therefore
00:08:46
their behaviour l. leaks if you want on the side of the person
00:08:49
who also tells us about the people where uh interested in understanding
00:08:54
now let's look quickly at the computer science part of the work
00:08:59
obviously we need to sense these interactions in one way or another this is an
00:09:03
endless uh search for the latest technology so what we have done in um
00:09:09
you know we're not over the years is essentially to use what has been state of the art at any given
00:09:13
time with respect to sense it so it would be typical technology the most of you are familiar with
00:09:18
a microphone arrays kind of an acceptable devices five years before like second to market
00:09:24
or actually neck devices that also and also to capture id be the worst colour information press also that
00:09:30
we have also use the other devices a better more more a specific
00:09:37
and i'm gonna show you just a couple couple examples of what technology can do with with respect
00:09:41
to the extraction of a specific behaviour out use one of them is a body motion
00:09:48
so what we have developed is i think the model of the body of also someone who said it
00:09:53
in the the idea of a modelling the body motion of the mardi gestures about person
00:09:58
is essentially to fit the visual observations this model and to see if the size
00:10:03
uh the position that is most likely that this person is gonna have
00:10:07
in the the world with respect to the emotions of h.
00:10:11
we're getting into the details where using this type of algorithms to be able to
00:10:15
understand whether a person for instance uh self touches which has some meaning uh
00:10:20
in a a from the perspective of communication or whether person crosses her arms
00:10:26
or with the person tends to just to give it a lot
00:10:28
this information will be used excuse for later analysis as a good example has to do with their face more
00:10:36
specifically with nodding as we know nodding is a strong signal to show that we're attentive to someone else
00:10:43
and you also can be used to to some big recording made the flower conversation so
00:10:48
what i'm showing there in that slide is essentially somewhat behind the reception desk
00:10:52
become the uh the algorithm has fitted the model of the of the face
00:10:57
and then when you see on the left top part of the slides essentially each time the
00:11:01
person not the client you can see this flirtation between up and down levels the signal
00:11:07
this can be extracted quite reliably and therefore you can estimate without without for instance
00:11:11
a person in the perception this is not a back to the customers
00:11:15
and if it looking back if it's not letting back we could use information later given feedback on why are you sending some
00:11:22
signal of a attention to your cost another example is a also quite interesting
00:11:29
which is gaze which is essentially to whom or where we're looking at
00:11:33
anything like is developed to the point in which we can actually identify between these two relatively subtle differences
00:11:39
cases when the person you know job interview is looking at the interviewer which is what you
00:11:44
have on the left side what cases with a person is looking away from the interviewer
00:11:49
this information can also be quantified and then used in the feedback a mechanic
00:11:55
finally we can integrate all this information using machine learning to try to make more high level inferences about the directors
00:12:02
and without getting into the because these shows how depending on the situation the performance of
00:12:06
the machine will be slightly more promising or less promising the problem is hard
00:12:11
and you can imagine what we're making progress with respect to a meeting or the inference of some of these quantities
00:12:18
it's important to say here that what we are ultimately trying to uh use the machine for
00:12:22
is to make the machine make some inferences about percy sub skills if we want not to application essentially there to
00:12:32
peer to peer applications from from these technology both of whom are meant to say to us to port
00:12:38
candidates the first one is how to write socially create filled uh feedbacks
00:12:45
or in in other words are behave us summaries that could be used by candy
00:12:49
they themselves might interpret seeing to getting that important interview for important presentation
00:12:54
after rehearsing you can have a desktop application or modification which you can actually see how you
00:13:01
behave with respect the physical quantities not just perception were actually physically quantities in one
00:13:06
of the challenges here is how information should be presented in the most effective way so
00:13:11
people can self reflect and use this kind of information as a tool for empowerment
00:13:17
the second one which has to do more with real time analysis is how we we actually if it back in real time
00:13:23
to for instance uh or a place in the hospitality industry about how they have performing
00:13:29
in on these we can base our work and which are about a
00:13:32
developing expertise has to do with given feedback right after the fact
00:13:37
clearly this is an example that i was outdated because he used to use the now ah
00:13:41
all that book last device but you can imagine application that you've you have time feedback
00:13:45
on smart watches on the phone so so uh candidates supply these are employees can actually
00:13:50
see how they performed in a more a discrete way than what is shown there
00:13:55
which is a a more engineering approach so let the completed just like pointing out to true challenges but it's actually
00:14:04
can be seen actually as all challenges because as you can imagine once we
00:14:08
have this technology one of the next step would be to scale up
00:14:12
the scale of for instance the collection of data online and in the afternoon one of
00:14:16
the start of with whom we are collaborating would be presenting some of these ideas
00:14:21
but one of the challenges is uh something that was my uh actually
00:14:25
mentioned this morning which is what happens with all this technology
00:14:29
when we train machines with data that he's not representative of the whole world
00:14:34
so this is a report that came out a few months ago
00:14:38
on a tool that amazon apparently uh develop and then decided
00:14:42
to discontinue because he was making highly biased uh decisions based
00:14:46
on uh information there was learn from their own employees
00:14:51
clearly there are some risks to the use of artificial intelligence of machine learning it would like to
00:14:56
what what is very important is that when it's to be very aware of what are the reasons why these systems fail
00:15:02
and what kind of false expectations companies or others can be
00:15:06
training in the population about the utility of these algorithms
00:15:10
what is clear right now in a technology world is that there are certain values that have not been developed consistently in
00:15:17
the development of the technology in this is to change in these values are often referred as furnace accountability and transparency
00:15:24
so we like to finish the talk by talking about each of these concepts
00:15:30
different it's a concept is that it's a very easy to understand but it's a difficult to implement machines
00:15:36
another word machine should be able to train enjoyable candidates the same chances and avoid discrimination biases
00:15:43
and when i was referring to new all challenge with this is still an old challenge for a a human resource
00:15:50
the second point is accountability which is that uh whoever develop this technology should be responsible for the parties
00:15:57
specifically i within should be all the double to know what's happening within the black box right now we're
00:16:02
still at a place where this is not satisfied as an industry in also some research community
00:16:07
and finally there's issues transparency which is that candidates or you know what people who uses technology should
00:16:13
know how machine decisions are made and how their video is used as part of this process
00:16:18
my take is that the technology won't be able to successfully get out of the lot and is all these constraints are satisfied
00:16:25
and uh with these uh like to invite you to uh discuss thank you first of all my collaborators interpret thank you
00:16:39
there is a lot of questions as you can expect i'm gonna try to go back on the slides
00:16:44
you know so you mention sonnets accountability transparency uh e. into to quote as it is today
00:16:51
there is very little of all of the above right um silence can discuss them in
00:16:58
the business model of many of the on like but formally use goes against
00:17:02
um these interests of standards accountability we see that a lot of silicon valley companies uh
00:17:09
i'm not very accountable maybe that becoming a little bit more transparency as you said it's a black box and now
00:17:15
there's even people who tell you like um google was noisy what's happening inside there
00:17:20
machine learning algorithm the body can like reverse engineer the reasoning the let this uh the to be sure to this person
00:17:27
your job what should do is really exciting and it's also very scary in some ways you can
00:17:32
think of the credit rating system in china for examples how do how can you make sure
00:17:37
but these values i applied to your technology at this point doing that the track record of the industry's not very good
00:17:42
yeah so i think i relaxation in the in the academic community developing machine learning algorithms
00:17:48
is how these values have been put on the side because what was uh the objective of the message have been something else
00:17:55
typically improve measures of performance were performance can be able to
00:17:59
be fine as making as many as fair as possible
00:18:03
but it is clear that technology conceiving that way has limits because all these concepts
00:18:08
are essentially put on the side so what they're gonna communities more aware and
00:18:12
starting to do step stores that is to actually design algorithms to have some
00:18:16
of these values embedded in the algorithms by design there is one part
00:18:21
another important part which is not so much better with technology but when is
00:18:25
is to be responsible and to collect the uh data that is diverse
00:18:30
that actually reflects the whole population you're supposed to be serving or working with
00:18:35
the issue of adversity statistical one because it goes beyond technology has
00:18:38
implications on just economic models it has implications on inequality
00:18:43
in ultimately justice so this is also an issue that was medically but needs to be being into the area
00:18:49
the age of transparency is an interesting one also because clearly there algorithms that i will
00:18:54
by definition are more transparent in other words easy to interpret what what's happening
00:18:59
but i think there's also some corporate responsibility when accepting that is actually good for the business to to just to hear
00:19:05
these values and i would that's what i would expect to see in the next season so it's a multi solution
00:19:10
uh in one of the aspects for for example the regulation to regulate what so it's which can uh and you
00:19:16
can the dukes uh another um question that can you know there's decided that uh for example old aeroplane pilots
00:19:22
uh i think that's young elephant by dots can notes by the planes anymore the main reason is that most of the time they put
00:19:28
the autopilot on so there is a diminishing of the skills when
00:19:32
you can rely on automation and you know in that system
00:19:36
i will come someone assistant tells me speak let's speak more but that means
00:19:39
maybe i'll stop actually looking at what's happening and i stopped thinking
00:19:44
yeah so something important here is a lot of looking for automation
00:19:49
and all right i see this as a tool that has to be used by person
00:19:53
as part of a process that is already in place in the the person who is using it says value on adding it
00:19:59
but i think the value of technology for instance for the real time system you could also think about it as
00:20:04
the covenant of that when activity tracking first came into into the uh the feel
00:20:09
so people found that that's an interesting gadget them put aside some of the people
00:20:12
decided to use it off and and gain some knowledge about the use
00:20:16
so the other princes runners or they are people who hate for six months or when you touch a certain goal
00:20:21
so but expect also the different types of users using the signal you forgive him raises any different ways
00:20:27
um this question like to use you know sure sure where basically you job interviews will be conducted by robots

Share this talk: 


Conference program

Introduction
Laurent Haug, Modérateur de la conférence
25 Jan. 2019 · 9:50 a.m.
262 views
Présentation des compétences R&D du site (4 instituts).
Laurent Sciboz , Directeur des instituts informatiques
25 Jan. 2019 · 9:51 a.m.
L'IMPACT DU NUMERIQUE SUR LE TRAVAIL
Paul Jacquin, RANDSTAD INNOVATION FUND
25 Jan. 2019 · 9:57 a.m.
LA "LOYCOCRACY" OU COMMENT GÉNÉRER L'INTELLIGENCE COLLECTIVE EN DISTRIBUANT LE POUVOIR
Christophe Barman et Elisabete Fernandes, Loyco
25 Jan. 2019 · 10:56 a.m.
COMPUTING FOR PEOPLE AT WORK
Daniel Gatica-Perez, Professeur à l’IDIAP Laboratory de l'EPFL
25 Jan. 2019 · 12:33 p.m.
Apprendre Demain
Anne-Dominique Salamin, HES-SO Valais-Wallis, Cyberlearn
25 Jan. 2019 · 2:05 p.m.
Comment la formation traditionnelle s’adapte pour répondre aux challenges de la digitalisation ?
Nicolas Debons, HES-SO Valais-Wallis, Institut Informatique de Gestion
25 Jan. 2019 · 2:11 p.m.
125 views
VIMA LINK
Raphaël Héraïef, VIMA LINK
25 Jan. 2019 · 2:16 p.m.
Projet de coworking à Grimentz pour les nomades numériques.
Fanny Caloz, SWISS ESCAPE
25 Jan. 2019 · 2:23 p.m.
ZOOM SUR LE RECRUTEMENT NUMÉRIQUE
Laetitia Kulak, GLOBAL HR TALENTS
25 Jan. 2019 · 2:29 p.m.
ÊTRE SPORTIF D'ÉLITE À L'ÈRE DIGITALE
Julien Lopez, FREERIDE WORLD CHAMPION
25 Jan. 2019 · 3:47 p.m.