Player is loading...

Embed

Copy embed code

Transcriptions

Note: this content has been automatically generated.
00:00:00
It will be here so I will stopped
00:00:04
immediately with the the basic scope.
00:00:08
So when we what we call the like normal
00:00:13
at a principle is that when we do some
00:00:17
movements okay okay okay I I think so
00:00:23
one label all or I would when we sign
00:00:26
you there are a margin behaviours that
00:00:29
are built in and there is like the
00:00:33
global features of of our of the
00:00:36
individual perfecting the movements. So
00:00:38
the log normality I will specify that a
00:00:41
little bit later on is kind of a
00:00:45
trademark of the perfect behaviour if
00:00:49
you are in perfect shape you will use
00:00:51
like normal to control your movement
00:00:54
and as a colour colour colour or Larry
00:00:59
young children for example will be
00:01:04
migrating toward like got normality.
00:01:07
And as we get older age and the L use
00:01:11
will move apart. So this will be the
00:01:14
key idea behind the talk five a panic I
00:01:20
I will talk interval time writing and
00:01:21
discuss other application later on but
00:01:25
we will define the handwriting strokes
00:01:28
I'll do we right will the the find the
00:01:32
concept of action plans. And is how we
00:01:36
can recover these will go back a little
00:01:39
bit too physiology to see if just makes
00:01:41
sense is and then we'll start showing a
00:01:44
list of application that will fit the
00:01:46
slide very rapidly on all the project
00:01:49
that are going on in this perspective.
00:01:52
And answer the question of are on this
00:01:54
if there are any so when we talk about
00:01:58
handwriting. So we're on the stroke is
00:02:02
kind of from the medallion in the on
00:02:05
writing that we are looking at and the
00:02:07
handwriting in general is a very
00:02:10
complex phenomenon the goal is to
00:02:12
communicate and and there is a message.
00:02:15
And there are many ways to analyse it
00:02:18
and at the very bottom of the of this
00:02:20
is the basic unit of movement a stroke
00:02:25
okay and similar things forcing it
00:02:28
servicing it's really when you sign you
00:02:30
want to be identified. But you are
00:02:32
using different steps. And the basic
00:02:37
unit is again the stroke and these
00:02:40
folks have been stardate for more than
00:02:43
one century. And they can be classified
00:02:46
a very simple strokes is like this the
00:02:49
stuff from what target and you want to
00:02:51
reach another target. And you've got
00:02:54
the villa simplified and there might be
00:02:55
some glitches at the beginning at the
00:02:57
end but globally speaking this kind of
00:03:00
isometric bell shave has been in the in
00:03:04
the air for one sensor a the that we
00:03:06
produce a isometric bell shaped just
00:03:08
like a kind mark of trademark of of a
00:03:14
fright primate batteries. So and when
00:03:18
we look at this yeah sorry whoops I
00:03:22
when we look at many of these so we
00:03:24
always see this kind of estimate rebel
00:03:26
saves that made very fast no this was
00:03:30
feedback almost rectilinear are and so
00:03:33
on. So these are the kind of properties
00:03:35
that are in there that in and this
00:03:39
strokes and the the I put this is that
00:03:43
they put for the laws that these
00:03:45
invariant property is reflect the
00:03:48
asymptotic behaviour of a complex
00:03:50
system made up of a large number of
00:03:52
copper neuromuscular network so we take
00:03:57
the brain and the arm as a black box
00:04:00
made up of many components which does
00:04:02
not know all they are connected. But we
00:04:05
know that they are couple. So and we
00:04:10
will use asymptotic there around too.
00:04:12
So that and this is the central limit
00:04:14
theorem work too. So basically the
00:04:19
model assume that I think given time T
00:04:22
zero O we give in order to move. This
00:04:25
propagates throughout the system to get
00:04:27
the velocity at the end and to get this
00:04:32
velocity we need an infinite number of
00:04:34
these box. And they are coupled with
00:04:37
does don't know but one condition is
00:04:39
that the company should be I guess the
00:04:42
order it propagates. And if I had the
00:04:45
chance to look at two places along the
00:04:47
that saying that I'm the curative time
00:04:50
delay at this point would be
00:04:52
proportional to the previous one and so
00:04:54
on so if there is a proportional the
00:04:56
relationship in timing of the go but in
00:04:59
the process you will get what we colour
00:05:01
like normal at the output no like
00:05:04
normal is a gosh and but with the
00:05:08
logarithm of the of the time here
00:05:10
instead of a time square what is
00:05:14
interesting about that equation is that
00:05:17
yes for example I I said make a stroke
00:05:20
what when there is a beep you perceive
00:05:25
the B you prepare a command that P zero
00:05:27
you gave it it propagates to your and
00:05:29
you start moving. And then you have the
00:05:32
like normal here that is characterised
00:05:34
by its time delay. And its response
00:05:37
time. So we have the surface under the
00:05:42
curve is the this and that you will
00:05:44
cover you is kind of them at the end of
00:05:47
the the the delay from T zero and sigma
00:05:51
here will be the an approximation of
00:05:54
the response time. So we have a few
00:05:56
parameters to describe the car. And so
00:05:59
when we want to make complex movement
00:06:02
what we do we are doing Victoria
00:06:04
summation of these like normal. So
00:06:07
we'll have a one moving towards is the
00:06:09
racial another one direction and we are
00:06:11
just something up and we got then some
00:06:15
of like normals the lord and I don't
00:06:18
one collation and then the the
00:06:20
difference interaction is described by
00:06:21
and you are a function the
00:06:23
interrogation on like normal and a
00:06:26
special case when we are doing just
00:06:29
appointing movement the fast pointing
00:06:30
movement then we defined the direction
00:06:33
the target. So there are muscles or was
00:06:36
your muscles moving towards the target.
00:06:39
We call them I got this. And the other
00:06:42
one is therefore breaking at the target
00:06:44
the antagonists. And then we have the
00:06:47
perfect a position. And the model is
00:06:51
then not the summation but the simple
00:06:54
subtraction. So to keywords to read
00:06:57
remember signal like normal model means
00:07:01
they carry all summation for complex
00:07:03
movement that are like normal model a
00:07:06
smaller model for pointing movement. So
00:07:11
the stroke is the ideal output of a
00:07:13
neuromuscular system reflecting the
00:07:15
it's impulse response which result from
00:07:19
the imaging be ever predict by the
00:07:20
central limit there so we have a formal
00:07:23
definition of what is a basic strokes
00:07:26
that is used in and writing and then
00:07:28
and then the signature yeah sorry. So
00:07:35
how do we rewrite then well we take
00:07:39
advantage of this like no no property
00:07:42
of the socks. We assume that in our
00:07:46
brain somewhere in the map we have we
00:07:49
have a virtual targets that are
00:07:52
indicating where we want to move. And
00:07:54
the system is sending commands to make
00:07:58
that victory all summation. And as we
00:08:01
get more and more used to making
00:08:04
movement smoothness emerge true bible
00:08:07
overlapping okay and we have the like
00:08:09
normal function of the for the lost a
00:08:12
profile which means that for writing
00:08:14
this letter he we have four fits well
00:08:18
targets from where we are given the
00:08:22
order to move toward a target. But
00:08:24
before reaching it. We just start at it
00:08:28
is sitting you know the other one and
00:08:30
so on. And so that's real strokes are
00:08:33
hidden in the signal that's a
00:08:34
challenging. We what what we get is the
00:08:38
blue curve and we have we want to
00:08:40
recover the hidden strokes in the
00:08:42
signal and so and the small cartoon
00:08:49
here right. So to do the right curve
00:08:54
once you have start moving their your
00:08:56
right already know that you reach it.
00:08:58
And you start moving it and with the
00:09:02
deal really kinematic terry we have all
00:09:04
the equation to the this and that is as
00:09:06
we can recover the end goal we can
00:09:08
recover the rusty and reconstruct
00:09:10
everything again the for those who have
00:09:13
not seen it I really enjoyed the small
00:09:17
five of everything is that if you
00:09:20
understand that everything is you can
00:09:23
even with so can we recover the action
00:09:27
plan okay this about this you know and
00:09:30
we did and we are the the answer is yes
00:09:32
not perfectly but yes for the does not
00:09:35
like normal equation we have a bunch
00:09:37
about software that uses produce at one
00:09:41
stroke what one that the like normal
00:09:44
movement and we extract. D zero. DUNC
00:09:48
about the I get is deviant signal of
00:09:50
the antagonist santa sitting factor
00:09:53
mean square or signal to noise ratio to
00:09:56
give the quality of their transfer
00:09:57
reconstruction for the stigmata normal
00:10:00
it's more complex we have nine of these
00:10:02
working and we know take all to to
00:10:05
reconstruct. And so which means that
00:10:08
when you write this type of patterns
00:10:11
and to use it if you can reproduce the
00:10:14
whole they'll say pattern with all the
00:10:16
solos and for each fraud you have all
00:10:17
the parameters. So from this word we
00:10:21
have T zero the when we your brain you
00:10:25
the although the sequence of order your
00:10:27
intention interval distance then all
00:10:30
your muscles reacts dynamo time delays
00:10:32
and time a response and that that are
00:10:36
the direction. And the direction of
00:10:38
each of these four a lot of information
00:10:41
so that men is softwares for that are
00:10:45
like non well we have also for optimal
00:10:49
sting like branch and bound to find the
00:10:51
exact solution many of these with with
00:10:55
prototype and so so we have tools they
00:10:59
are not perfect but we have tools and
00:11:01
these tools are allows ask for example
00:11:03
for that to like normal to from the
00:11:05
blue curve we are able to extract the I
00:11:08
gonna and the antagonist movement. And
00:11:12
analyse their properties okay for the
00:11:15
more complex sigma like normal again we
00:11:18
are X build extract the target. And
00:11:21
repose use all the movement almost
00:11:22
perfectly and each time the parameters
00:11:26
continues used to study any phenomenon
00:11:30
we are interested in and that's what so
00:11:34
we are able to recover the hidden salt
00:11:37
in the signal and we are still going
00:11:42
all we can optimise these type of
00:11:44
things we have a power corporation here
00:11:46
with the people in violence yeah well
00:11:48
we have been able with a more
00:11:50
exhaustive search to reviews the number
00:11:53
of log normal to get the better signal
00:11:55
to noise ratio for there is room for
00:11:57
improvement but with what we have we
00:11:59
can already do many things and what
00:12:01
that's what I was so you in the the
00:12:04
next minutes. So but first is it just
00:12:08
an answering dream or is it the meeting
00:12:11
meaningful there are two things that we
00:12:13
can test or the very first one is that
00:12:16
we are pretty thing that we are forty
00:12:19
Chantal and evolve there is a time in
00:12:23
our brain where we have the given the
00:12:25
order. So if we do. EG and ask people
00:12:31
to do that are like normal strokes. We
00:12:34
predict that there will be an evoke
00:12:36
button so ERP buttons so it should
00:12:40
exist and it should happens in this
00:12:42
around the T zero. So we did the
00:12:45
experiment the narrow scientists
00:12:48
collected the data out this was sixty
00:12:49
four electrodes do their business with
00:12:52
all this tools and they say yeah yes
00:12:55
there is a and ERP we take the velocity
00:12:57
profile extract reconstructing that is
00:13:00
zeros with the the mean the and the
00:13:03
scene the the woodwind though the they
00:13:06
are and what you zero this in the
00:13:10
window of or DRP abundance not we
00:13:15
didn't expect is that the actual that's
00:13:17
where the most sense as they were on
00:13:20
the left hand side of the brain and
00:13:21
other people are right handed. But this
00:13:24
I would not put my head on the on on
00:13:28
this we had only have twelve or fifteen
00:13:31
subjects and it was not really a
00:13:34
location you know it it up and it might
00:13:37
be just by luck that say that the zero
00:13:40
it means something the other right what
00:13:42
is just is the from the famous
00:13:44
cumulative time they right. So what we
00:13:47
did we did EMG analysis we started last
00:13:52
people do you suppose and we have seven
00:13:54
muscle studied here. And here again we
00:13:57
the the the the people in the MG the
00:13:59
the analysis of these signals we did
00:14:02
the others as of the last day they do
00:14:05
the the the technique the basic
00:14:07
techniques to extract the envelope and
00:14:09
so on and we end up with the bar almost
00:14:11
perfect proportional relationships with
00:14:13
all all pairs in the Shane so showing
00:14:16
that there are there are proportional
00:14:18
effects in the timing all this
00:14:21
subunits. So to to point that give us
00:14:25
some confidence that we are maybe
00:14:28
touching something that is meaningful
00:14:29
also for buttons computations and
00:14:34
finally on that there were still a
00:14:35
comparison we started well the like
00:14:38
normal stand with respect to all the
00:14:40
others are many other models are we
00:14:42
don't claim that we are the only one
00:14:44
and for some application different
00:14:45
other models might be better than us.
00:14:48
But what we know is that the like
00:14:50
normal is the ultimate one if you get
00:14:54
the fashion all these model loss
00:14:58
suboptimal. And if you look at but gosh
00:15:02
and you are neglecting the dependency
00:15:04
in your system which is not
00:15:06
biologically meaningful because we know
00:15:10
this is a within the same electrolytes
00:15:12
and things like that it cannot this
00:15:14
scene as independence of the like
00:15:16
normally yeah so they were the only an
00:15:20
experimental a there is something
00:15:22
insensitive to work with like novels.
00:15:25
So where do we go from here as your
00:15:27
back to like normal it they will just
00:15:29
show that we have observed this move
00:15:32
very briefly and then what might
00:15:34
interest you more what we are doing now
00:15:37
with this. So back to the long on
00:15:41
Monday what we did we learning we look
00:15:44
simple and we have them to do some
00:15:46
various curve and so on. Uh and we show
00:15:51
that we were able although the model
00:15:52
from the south was design for rapid
00:15:55
movement it works also for slow
00:15:56
movement because we are something up
00:15:58
like normal then so we are able to
00:16:00
reconstruct various type of the luster
00:16:02
profiles and trajectories yeah and yeah
00:16:06
and for what is pretty what we pretty I
00:16:09
guess that as the students or the kids
00:16:12
are getting more and more in the
00:16:14
controller there nor almost or is it
00:16:16
the the number of like normal to
00:16:18
reconstruct is really is reduced. And
00:16:22
the fitting is better and by the way
00:16:25
the signal to noise ratio in the
00:16:27
reconstruction the number of like
00:16:29
normal then the race so these are three
00:16:31
indicators of the you're like normally
00:16:34
this status at least we have kids from
00:16:37
three four to five years old and we
00:16:38
were able to see where they stand in
00:16:42
the evolution of a tool and emigration
00:16:45
to or like normality ageing we should
00:16:49
move away. So we are again the one
00:16:52
hundred percent with some with dress
00:16:54
rates for Chris factor others with no
00:16:57
rain so and again the delta like normal
00:16:59
test the signal to noise ratio so the
00:17:02
quality of the construction decrease
00:17:04
with age. And if you have a risk
00:17:06
factors or less aide this'll be is
00:17:09
steeper the same thing for triangle
00:17:13
drying is so sick like normal of the
00:17:15
other one was that the like normal test
00:17:17
again these three indicators here shows
00:17:20
that we I'm good predictors so it looks
00:17:29
as if we are as a baby when we do all
00:17:32
these bubbles we are just collecting
00:17:34
and synchronising our muscles the
00:17:38
they're all system to finally develop
00:17:41
like normals and then we learn to use
00:17:43
like normal to concatenate them some
00:17:45
them to write down and number movement
00:17:48
and at the very end we will move away
00:17:52
from the like normal today. So what we
00:17:57
do next what will so five flip the
00:17:59
slides on various three things E
00:18:02
security. So signature verification
00:18:04
with people in other yet that we are
00:18:07
studying ageing which not allowed a
00:18:11
long time span but again we can see the
00:18:14
same behaviour of the number number of
00:18:17
like normal going down. And the signal
00:18:21
to noise is so going up you're going
00:18:22
down again and number won't know about
00:18:24
going up sorry say well they county
00:18:27
ageing but I think it's more adaptation
00:18:29
to to to the new with do these were
00:18:31
done on tablets and I've phone and
00:18:34
people were not used to sign but anyway
00:18:37
with the people literally we are
00:18:39
studying the stability rinsing answer
00:18:41
okay so what is interesting also is to
00:18:43
find a way to see if you are stable or
00:18:46
not then you know saying it or if we
00:18:47
could have kind of index of the
00:18:49
stability they then when you sighing
00:18:52
and you want to be checked depending on
00:18:55
your stability index the system could
00:18:56
have kind of personalised threshold to
00:18:59
accept or reject you so with the my
00:19:01
friend use of the film oh we are rising
00:19:04
and studying and quite to quantify the
00:19:06
stability of various signers depending
00:19:09
on the way we control the like normals
00:19:11
we also have a another problem in
00:19:15
signature verification we don't have
00:19:17
enough signatures if you want to do
00:19:20
handwriting recognition there are your
00:19:22
database with millions and millions of
00:19:24
characters or words and so on the
00:19:26
signatures it's a proprietary
00:19:28
information people don't want to share
00:19:30
it and so on so we decided that with a
00:19:33
small database we could generate. So
00:19:36
what we do with a guessing it's or the
00:19:38
the reverse signal like no why
00:19:40
selection we have the power meters but
00:19:42
the noise on these parameters and
00:19:44
generate saying it's there's not much
00:19:45
as much as you want that there's all
00:19:48
your job we have also yes we have also
00:19:56
use the signal like normal parameters
00:19:58
to kind of qualified a forgery and
00:20:03
waited the those thresholds for
00:20:06
accepting or rejecting a signature
00:20:08
based on the quality of the stigmata
00:20:11
normal like because of fort rose will
00:20:13
not reproduce perfectly they will have
00:20:15
more like normal they would and the
00:20:18
lower signal to noise ratio and so on
00:20:20
so we are able to wait the decision
00:20:22
here based on this model too but we
00:20:25
also use a brighter listing them out
00:20:27
and almost directly by reconstructing
00:20:29
and looking at this time. And again we
00:20:32
have been able to and achieve the state
00:20:36
of the art result for skill forgeries
00:20:39
just using this model we do rubber
00:20:43
score normalisation days again notes on
00:20:46
a similarly normal model well yes the
00:20:51
it will be but some it's next week
00:20:53
design your system with on the one
00:20:57
thing it's or so you you have supports
00:21:00
on the side one time and we'll build a
00:21:02
system directly whoops what of them
00:21:08
something yeah so we are able we take
00:21:16
once again sure excited parameters on
00:21:19
reprise use as much as we want to learn
00:21:22
all the sign you're designing. And the
00:21:24
pro the the the results are pretty good
00:21:26
not state of the art yet but with only
00:21:28
one thing if caps are generation again
00:21:33
the we you know capsize all of these
00:21:35
buttons that we have to recognise
00:21:36
distorted button the computers are the
00:21:40
button on the image recognition are
00:21:42
getting better and better so the the
00:21:43
gap between human and computer is
00:21:46
diminishing so handwriting is more
00:21:49
difficult to recognise and we generate
00:21:52
unwritten kept capsize and add noise
00:21:55
and all the things and then we increase
00:21:58
the gap in some condition here the
00:22:00
computer almost zero recognition and
00:22:03
the manner in the eighty five or ninety
00:22:05
percent he helped this is another one
00:22:11
one the idea is to this is being done
00:22:14
with the people in basra. We want to do
00:22:17
word spotting what people are writing
00:22:19
and each time you you were the words
00:22:21
speed for example we send it on the
00:22:24
cloud and the D analysis and keep track
00:22:27
of your parameters. So that we can see
00:22:30
if you are still in the old log normal
00:22:33
nor normally they be if you're or if
00:22:35
you're moving away. And so maybe if you
00:22:38
are moving away there is some problem
00:22:41
coming and we just don't know it's not
00:22:43
the specific to this kind of a balance
00:22:46
Arthur meter we have we have that
00:22:49
thermometer but the scale is not
00:22:50
defined yet well and we have a
00:22:54
prototype working on the on the
00:22:56
something tablet on this another star
00:22:59
days other start is with browser race
00:23:02
or crest factor packing some disease
00:23:04
alzheimer's disease. Because according
00:23:07
to the like normal to principal all
00:23:09
these people should move the way and
00:23:12
maybe we can see something in this. So
00:23:14
back to the same study. We have shown
00:23:17
that there is a relative definition the
00:23:19
relationship between the presence of
00:23:21
focus factor and the characteristic of
00:23:23
human movement as an allies with
00:23:25
symbolic normal parking well and we
00:23:30
repeat that would various type of task
00:23:32
here is an oscillation task which come
00:23:34
from the previous results also biking
00:23:38
some disease. When people are writing
00:23:41
this is a all all percent. So it's not
00:23:45
perfect recognition more target here
00:23:47
you can see then on on the one that we
00:23:49
would expect and this certain number of
00:23:53
like normal as compared to parkinson's
00:23:55
on writing where than the number is
00:23:57
increasing and a lot of targets and a
00:24:00
station so again. It's then the signal
00:24:05
to noise ratio number of like normals
00:24:06
are indicator of the but it is we have
00:24:10
kind of a metric but it is not fall a
00:24:15
chronic quantified yet we are doing the
00:24:19
same type of study in Montreal the
00:24:22
Montreal dramatic concert you where
00:24:24
they are asking people from without
00:24:27
parking sound done the control to do
00:24:29
biking. And just the ipod is that just
00:24:33
doing something the size every they
00:24:35
will improve your own globally if
00:24:37
you're so we ask people to do that
00:24:40
online normal strokes and the the at
00:24:42
the end then we follow them every two
00:24:45
months for almost almost a year now.
00:24:47
And then again it's it's it's
00:24:50
alzheimer's disease to study one in
00:24:52
Italy they have built up a database of
00:24:56
and written words mama that we can
00:24:59
reconstruct lee and and we are
00:25:01
realising another one in Paris because
00:25:05
this is for handwriting but this this
00:25:07
is also right for the role of the
00:25:09
angular velocity this is right for the
00:25:11
the the rest the had a high we are that
00:25:14
are like normal build a private. So I
00:25:18
will not we can reproduce eye movement
00:25:20
but what is maybe not well known is
00:25:23
that and when you have a disease like
00:25:27
alzheimer's and this not as clean and
00:25:30
we can reproduce it but you can then
00:25:32
see that there is more like normal that
00:25:34
is the the reconstruction is as good.
00:25:36
And there is a test in that I didn't
00:25:39
know before starting that they ask
00:25:41
people to look at that point on the
00:25:44
screen and the the maker across and
00:25:47
have people to move their eyes yeah and
00:25:50
if you are healthy you are kind of
00:25:53
regrouping your target almost perfectly
00:25:57
and when you have alzheimer's you on
00:25:59
less worn out of the store like that.
00:26:01
And what is interesting now is that
00:26:03
worry about twos for each of the eye
00:26:05
movement we are able to extract the
00:26:07
parameters now the parameters T zero
00:26:10
and D reflect the end goal of moving
00:26:13
their eyes and you and Seymour reflect
00:26:16
them the quality of your muscles under
00:26:18
the control of your muscles. So still
00:26:21
again on their starting just so
00:26:24
analysis with people in there and that
00:26:26
I've that last week. They are
00:26:28
interested in designing tablet with
00:26:32
just your commands. But they would like
00:26:35
to minimise the number of just so that
00:26:37
you have to give to the tablet before
00:26:39
it recognise you you know you want to
00:26:41
make crossover to erase if I ask you to
00:26:45
make twenty crosses to erase is a you
00:26:49
will draw the tablet if I ask you two
00:26:51
or three but are we construct one day
00:26:53
from your own you're not bothered and
00:26:56
that's what we did we regenerate
00:26:59
sensitive gesture. And what we saw that
00:27:01
the system was learning faster one one
00:27:06
yes and the other one is that when you
00:27:08
decide to change or add you gesture
00:27:12
instead of losing that memory or the
00:27:15
the the the the preservation is smarter
00:27:17
to so two for the price of one and we
00:27:22
are doing the same thing at EDS in
00:27:24
Montreal with them I mentioned really
00:27:25
and then again in that and so we have
00:27:28
this I website where people can design
00:27:31
their own gesture and then regenerate
00:27:36
whatever they want interactivity to
00:27:40
define whatever they want with the
00:27:42
jester you user interface. And
00:27:45
difficult to see which one is real
00:27:47
which one is sent to yeah in in yeah we
00:27:52
are again the same problem they would
00:27:54
like to reckon recognise Bangor they
00:27:57
don't have enough space Simmons so
00:27:59
again regenerate database of band hear
00:28:02
the problem was that there are many pen
00:28:04
ups. And we have to recover and do not
00:28:07
create discontinue is in between each.
00:28:10
We solve this problem and we are able
00:28:12
to reconstruct. And we that S which
00:28:14
Israel which is that they take the real
00:28:17
are here it was not undersized. But
00:28:20
they had to find which one was it which
00:28:22
one then again fifty percent chance
00:28:24
they could not recognise. This allows
00:28:27
us to define the range of parameter out
00:28:29
of how much we can distort to to keep
00:28:34
you meant like minutes you learning
00:28:36
with kids again in by the loop and also
00:28:40
you know and we have to we give them
00:28:43
value we have three four and five years
00:28:45
all you had to do some past like this
00:28:49
handwriting. And again we were able to
00:28:53
show where the the number of like
00:28:55
normal signals minus one so on were
00:28:58
able to qualify the multiplicity of the
00:29:00
kids. They go here what is you know
00:29:03
when we start at school we we have to
00:29:05
succeed then if you want to say is not
00:29:08
info perfect control you are four years
00:29:11
old in in what was it a in the five
00:29:14
years old school if if it can be
00:29:17
detected you will need more help and
00:29:20
you will not be affected at the reverse
00:29:22
if you are the forty years for and the
00:29:25
with the motor still five years maybe
00:29:27
we can give you more so this is the
00:29:29
type of thing that we are looking at
00:29:30
yeah and we have also done that with
00:29:35
younger doing just crippling that two
00:29:38
years the they are already using like
00:29:41
normal without knowing it and we start
00:29:45
a study on the EDH the hyperactivity
00:29:50
syndrome is in Montreal with kids again
00:29:53
to try to see as compared to a control
00:29:55
what is I think I have the and then all
00:30:00
the wrong conclusion so the idea would
00:30:04
be here that your your kids want to
00:30:07
play football before up there again you
00:30:11
will make then strokes and you go and
00:30:13
if you as a commotion the you but yeah
00:30:16
the commotion do your so again out out
00:30:19
the differences in the in the
00:30:20
parameters and if there are maybe
00:30:22
you're having some problems non
00:30:25
invasive task you know what's as
00:30:26
important here. You just need a tablet
00:30:28
and you do things no no blood no
00:30:32
nothing to disturb you no buttons
00:30:37
already have done the same thing okay
00:30:39
and I think that just the tip of the
00:30:41
iceberg we can do more than that
00:30:43
because well we have seen online and
00:30:48
writing recognition see answer
00:30:50
education because we have a new
00:30:52
representation space and we have the
00:30:54
automatic segmentation two key features
00:30:58
of this approach. So we can analyse the
00:31:02
writer's style we can you design you
00:31:04
online recogniser tools to help
00:31:07
children to learn and writing this is
00:31:09
what we are starting with people in the
00:31:12
in the then they are making a huge
00:31:14
collection of the data thousands of
00:31:17
kids every mons to they've developed
00:31:20
learning lessons and we want to have
00:31:22
access to this to help them on it or
00:31:25
the evolution of kids in biomedical
00:31:29
signal processing we have a new set of
00:31:31
parameters three describing the brain
00:31:35
for describing the the arm so it's
00:31:38
pretty interesting. And then we can
00:31:40
design oops I commanded S for for
00:31:43
different the illness or destined tools
00:31:47
we can also begin to more draw a called
00:31:49
and whatever it things and we have a
00:31:51
new tools to study these it it is nuts
00:31:55
replacing anything in just adding up
00:31:58
new way and you wind up aside. And I'm
00:32:02
here for the open fields and with my
00:32:04
friends of I will we know that these
00:32:07
are the smooth the go model like that
00:32:09
we can do with like normal so on top of
00:32:11
a muffin come design protesters taken
00:32:14
so on. You mentor with models. And we
00:32:16
will if we stop things like that we
00:32:18
would be whoops nice enough yeah yeah
00:32:23
really don't want me yeah yeah oh it
00:32:57
works what a movement on the big make
00:33:00
screen like this. The parameter are
00:33:02
different larger the dog larger un
00:33:05
signal because the console are
00:33:06
different with the friends in the in
00:33:12
the long run and the with so I also we
00:33:15
start the start studying rhapsody
00:33:17
making and know that think Daniel is
00:33:20
here so they all are interested in
00:33:23
reproducing in fine arts like this. And
00:33:27
they use the sigma like normal model
00:33:28
with some tricks to generate nice
00:33:32
patterns like this and having some
00:33:33
robots design you the and this is what
00:33:37
we are going to discuss that in
00:33:40
tomorrow with them. And also with us in
00:33:43
my there is a nice video here that you
00:33:46
can have a look at where the robots
00:33:49
Baxter is signing and learning to write
00:33:51
and this is kind of inspired from what
00:33:54
we are doing but there is still a lot
00:33:56
of things to do finally speech. This is
00:34:01
the second paper that will present on
00:34:04
this with the people in the last but
00:34:09
alas and yeah and generally had might
00:34:13
be obliged to go to London every in
00:34:15
winter next week next year. So we we
00:34:21
they wanted to say well we have this
00:34:23
mothers working pretty well on
00:34:24
analysing why can't we apply it to
00:34:27
speech. And the idea is that we use
00:34:30
that to form and and with the file in
00:34:34
your presentations space formant one
00:34:36
informants to rip like X and Y and of
00:34:39
all are the visceral target and we are
00:34:42
moving on one bald another by using
00:34:45
consonant. And we define a velocity in
00:34:49
that space. And it up on that this the
00:34:51
loss that is perfect yeah to you can
00:34:57
reproduce exactly almost exactly which
00:34:59
signal to noise ratio very high the
00:35:01
speech and now again what we are
00:35:06
interested in is like recently these
00:35:08
and getting the rose there are so many
00:35:11
lower inject colour speech and we have
00:35:14
the we were present in a couple of
00:35:16
months and in in Spain yesterday where
00:35:19
we show that not only the parameter the
00:35:21
symbolic normal also are able to
00:35:24
characterise but there are different
00:35:28
behaviours. I think that signal buys
00:35:30
crazy here decreasing there and so on
00:35:32
so or interesting patterns that are
00:35:34
emerging Virginia I guess the signal
00:35:36
the in that conference yes in so these
00:35:42
are the results of national
00:35:43
international collaboration with plenty
00:35:45
of people that I thanks I previous
00:35:47
supported it like other them. And help
00:35:51
that will do the same here with us
00:35:53
ceiling the I think it's time for
00:35:57
questions. Thank you very much for your
00:35:59
attention I guess so yes yeah at that
00:36:19
age. So strikes is strokes. It's
00:36:36
deviating from two types of the
00:36:42
deviation one is on the it's maybe it's
00:36:47
intuitive at the very first thing is
00:36:49
you still have remote normal but you
00:36:51
use more of these two to do your task.
00:36:54
And later on you know you're not like
00:36:57
normal you're moving from the fitting
00:36:59
that so the number of like not Moses
00:37:01
one indicator. So you you are using
00:37:04
more but then the signal to noise ratio
00:37:07
the quality of the reconstruction of
00:37:09
each of these like normal he's also
00:37:10
decreasing so then it means if we take
00:37:13
into about infinity of a subsystem in
00:37:16
terms of the timing better timing in
00:37:19
the in the proportionality this is
00:37:21
probably decreasing slowly as we agent
00:37:25
help problem. Uh so but this is a
00:37:28
global model you know this this is a
00:37:30
kind of a general indicator then we we
00:37:34
we'll to say that well once you added
00:37:35
that maybe got your medical doctor and
00:37:38
they can give you some well that we
00:37:41
need something more test data for a
00:37:43
while but this is kind of a preventive
00:37:45
tool that we expect we'll have I don't
00:37:47
know if I answer your question so it's
00:37:49
you're right the boat both not two
00:37:52
phenomenon that are appearing. And I
00:37:55
think my feeling from the the the study
00:37:59
that we add with normal age people do
00:38:01
that we are kind of protecting
00:38:05
ourselves so instead of doing that you
00:38:06
know and we will do its motor it takes
00:38:11
more like normal and we don't break our
00:38:13
arm at the end nothing like that. But
00:38:15
if a kind of a international protection
00:38:17
okay yeah yeah well yes yes both when
00:39:10
we are interested in duplicating and
00:39:12
having a large number we take one
00:39:15
target signatures and then we generate
00:39:17
many of these if we want more valuable
00:39:19
they we can five can take five
00:39:22
depending on the top for example one so
00:39:23
they will want to have one the one
00:39:25
thing actually so we did everything
00:39:27
from one thing answer. But and also on
00:39:31
the recognition if I have ten of your
00:39:33
seeing it so then I reckon slide and
00:39:35
then I can have a look at the
00:39:37
distribution of your parameters and
00:39:39
maybe it give us more information about
00:39:41
the way you sign and if I'm trying to
00:39:43
imitate you I will be out of drag
00:39:45
because even if the image is nice. I
00:39:49
will not be in the range of parameters
00:39:51
so bowl. We can play with one yeah and
00:39:59
we that moron handwriting seeing
00:40:01
assures it also said people don't want
00:40:04
to give money seeing it so that no one
00:40:06
except force days and we when we got
00:40:08
fast forty times a person to sign you
00:40:11
know it kind of boring so they don't
00:40:13
want to and then a are are going with
00:40:18
you know with the learning and all
00:40:19
these things it need you views about
00:40:21
the data now so if you really want to
00:40:24
exploit that in handwriting and
00:40:26
probably in speech also what what we
00:40:28
see here. I see a lot of the future
00:40:31
application on speech I think with the
00:40:33
same kind of model. And eye movement
00:40:36
that would be everything was totally.

Share this talk: 


Recommended talks

Towards face-to-face conversations with social robots
Joakim Gustafson, Associate Prof. , Royal Institute of Tech. , SWE
June 21, 2012 · 9:15 a.m.
Eyeware | solutions de suivi du regard
Jean-Marc Odobez, Idiap
Sept. 8, 2021 · 9:27 a.m.