Embed code
so uh i'm had everybody i'm sorry from matrices or building
basically from here um so first of all uh what is survey
select from your fee and white is used for and prosthetics
so in recent years we have seen the more and more advancements in a a technology in portable technology
yeah it is for seen that in the next year is
the solution we live translates to robotics as well
and to a human capabilities enhancement also for impaired people
however this field require the capability to the record signals from human beings
which uh and uh that's one euro can be useful for these
so what is selection menu buffy lets me your f.
is the measurement of electrical activity decrease muscle contractions
so when we thing to do a movement the movement originates into the brain any travels to target muscle
various are sexy we have electric chemical changes that leads to the
a muscle contraction that it produces an action and action potential
this action potential can be detected put in some lectures on this phase of the screen
why's election your of important for interest headaches well simply because i met
the muscles that control the can't actually a locating into before
so people that are implicated below the elbow they still had the muscles that used to control the hand before okay
uh lets me out of the s. b. used in hand prosthetics seems the sixties
uh nowadays uh eat uh has several pros and cons so nowadays
we have to make any uh i'd hands everybody can suffer very very advanced
lay mechanical point of view that can rotate the time and the released
that can perform up to thirty c. six a a program movements different movements of
the hand different press however they have also several counts so would be
a control c. start extremely rudimentary the training times along because
they require dedication of the subject to control the prophecies
and the uh the the alternative control strategies that have been developed are not
extremely usable nevertheless be a first commercial part
recognition systems have been recently um
released what is essentially take state of the art in hand prosthetics
um many times have been done to the control robotic and we've
gotta recognition movement uh classification techniques with proportional instantaneous control
uh with multimodal data fusion including taxidermy tourist pressure sensors or computer sensors
actually there are also some more complex um uh strategies that
in boulder for example targeting master innovation to translate the
uh narrows to control other mascot of the muscles um
that are not used for x. p. c. tasks
or a a neural interface is that contrast for the
sensation reader related to the shape and be a
stiffness of the object to object is objects to the uh to the person that using the processes
the most common approach is a a incentive researches however
to perform pattern recognition for movement classification would see
and the however in general most all all the c. stands
rely anyway on a similar hardware acquisition set up
however software setting the stage so stuffy selects from your
record a lexus are ways around before are here
the it's a lecture me over the signals heels
recorded it is proposes amplified filters sample
the segmentations perform feature extraction perform and uh finally the movement is classified for air
regression can be applied rather take me some of these can be applied
uh again into the state of the art uh we have a um
many publications actually so homely here that uh publications
uh say here till two thousand eleven or more than forty
and issue the deck solos control seem to be reliable
however uh the data ah really of
those publications is not publicly available
the acquisition methods are encouraging us or d. r. can be very diverse in different
publications does the results cannot be compared uh and reproduce the you see
so what he's uh um uh our experience he know electron
biography data sharing in at the state of the art
well i recognise that the uh uh data sharing intervals from your fee is that can give an excellent impact
and can be an excellent source of collaboration and that it can improve while the knowledge in the field
so it has been a shown in the the church aren't described in previous presentations
that uh a publicly available data can allow to improve different fields
uh includes come including computer vision medically machine and many others
however till two thousand eleven interface selector my your feet
did not have any publicly available benchmark data sets
so what we did into the state of the art that i described before
with so many publications that was what had happened to thousand leaven
was to create and refine are produced by the position protocol for a lecture me over three
profits control to up to record a publicly available data base and to realise it
you know the two uh allow what why tess and compare is on off classification algorithms you
know the to finally of mental dexterity also faces for me over the process is
and to reduce the uh training times so to make better prosodics that
these projects was uh uh funded by this this national transformations actually
it was led by each yep but by backup put up
any to this ad as part has a us and the
german or space centre so what was that the um
the acquisition protocol uh all the the of these database
roughly it consisted of having d. both intact and and ability subjects
uh uh watching some movies of movements on the screen of the laptop
we had several sensor located on the forum of the subjects so for amputees we add
uh electronic your having a minute censored that and uh
uh accelerometer is point of subjects we also had a separate
lot to record the ground truth of finger movements
we recorded the letterman europe the signals and we are from
the analysis in order to allow a robotic process control
those are the movements that we recorded so we
have a first set of uh um
for a set of single finger movements a set of basic
and increased movements so for example with quotation with
to be nation and a set of a grass and functional movements and also a set of course better
the movements where chosen by a d. a hand
taxonomy and hand remote expenditure sure careful
um the uh results so would be you know pro database actually obviously the first one
and deceitful because in the field but actually rather continues increase in it and it's
get in more more use it is publicly available it includes a six data sets currently
three of them are regularly is that was the other ones are waiting for a publication basically for peer review that it's quite a long process
it currently includes ninety seven index subjects and eleven trance radio i'm pretty subjects
performing approximately fifty different movements and uh uh currently we are record in
a a more and more data sets with different positions adopts
um and the database current have more than one hundred and fifty users
it's like two three official publications describe in
the database that got a uh
approximately a one hundred and fifty citation since two thousand and twelve
but obviously it's led also to several other publications related to the analysis of the results
it is continues an increase in the number of publications related treat because finally data are useful and
a number of citations because people find it useful and the easiest lead into a um
increasing number of international collaborations because finally people are contracting out
maybe to just to to recordings in the same way
uh it's allowed as results to test a a different data no these methods
so traditional classification metals deep learning methods regression with
the mission will then i'm ten working
we're paying a a way to show the possibility to uh recognise up to eleven different
hand movements without us a higher than ninety percent in and i'm pretty so
a a set of movements that could be used in real life settings let's say
um and the utility show the relationship between clinical parameters
and movements classification accuracy in a hand amputees
so for example we showed that body mass index uh
uh and that s. body mass index increase
the classification accuracy decreases because basically fat x. x. as insolent
we showed that the the longer the uh for our percentage remaining
big better would be the accuracy with which we can record
be uh we we can classify different movement which
is quite obvious but was not demonstrated before
we show that actually classification accuracy increases as
fun tumbling sensation increases so funny
sensation is the sensation of having steal your hand here once you have
been updated and we show that the sensation the higher the sensation be higher would be actually
it's which we can uh recognise the uh different movements and we show that actually
the longer the time last night imputation the high there
would be the accuracy which is a bit we're
and that but but is consistent somehow with the um
or yes we horace consistent with some results
a related to to to how be our cortex
adapt to the changes into the body
so what is our experience finally in a a lecture me over fee reproduce ability
or experiences that reproduce it really is difficult in this
field many parameters can strongly influence the results
it is fundamental to explain in detail data acquisition
and analysis procedures and to pay extreme
attention actually went reproduce in the result so when sane but you're reproducing the results
and it is fundamental to have benchmark data sets and proper comparison massive
so uh you someplace else mayor forehand prosthetics actually
it is fundamental to putting results that can be translated
to real life applications because we want amputees
wow one day hopefully still want to be able to control robotic and
as in science fiction movies uh the so depreciation setup but
can a influenced strongly reproduce immediately we tested in uh our database
uh for different acquisition setup ranging from a few hundred euros
to almost twenty thousand euros and uh uh um
actually with steamed that uh uh the classification
results are all ten comparable considering be
a senior position setup even considering the which devices and very expensive devices
this results uh are we going to do to to be published actually we
as uh since we aim to the uh um to translate a results to real life and um
we should we see this scene that acquisition procedure is can strongly influence reproduce ability as well
so for example training in uh and testing on a different
date acquisitions made we did seem acquisition setup but simply
pull it needs out and put in it again on
to the subjects can uh affects strongly classification accuracy
we have seen obviously as we all know that different
uh uh data analysis methods can strongly influence
uh the uh the the in the results of classification
and particular here we showed it this summer that
she computation or networks apply to some faceless from your face basically get
normal results actually so the d. d. didn't improve the results
that much but obviously we want to uh to to
the performance to continue analysing these the application of these uh two techniques to this field
we have seen that actually the number of
classes can uh strongly influence the results
so uh actually because the chance level varies with the number class
in classes and considering a specific data set feature classifier
the classification assurances theoretically expected to decrease as a number classes increases
the uh and we've seen as well that class
balancing then that's strongly influence reproduce ability
and uh so for example if but we have unbalanced classification accuracy
depending on the data set obviously but uh in several in some data sets
i mean i'm balance specification actually can be up to
fifteen percent higher then balanced classification at u.
s. c. d.'s is basically due to the fact of specific classes that uh are classified with
accuracy that are much higher than the other for example you know lecture my your perfect rest
and this is usually not pointed out in a most of the church
so finally in conclusion and i'll time uh the uh um a survey sites from
your work beforehand prosthetics state of the art so the scientific results that
have been shown our encouraging but the the are not easily comparable one to
the other they're not robust enough for real live requirements as well
we got a letter from your every data sharing so we really used
be biggest database and of first immediately routes you're the only one
database for surveys that american data counted include more than one to
one hundred subjects perform in approximately fifty different hand movements
he's allowed to improve public knowledge knowledge in the field and it allowed to impart
descended literature and to be a a and became a source of collaboration
we're gonna lecture me reproduce ability uh well uh we sold that many parameters uh can strongly
influence results so as the all within before we
have a uh it's difficult to reproduce
results uh well it is fundamental to explain the us in a publication i think to explain the
tape the data acquisition and all these procedures and to be extreme attention when reproducing results
and of others and it is fundamental as well to have benchmark
data sets improper comparison methods in order to allow up
and proper advancement in the feel that it's not like saying okay i got two percent
more get these five percent more but on what and the with which met
okay so thanks a lot for your attention

Share this talk: 


Conference program

Welcome
Sébastien Marcel, Senior Researcher, IDIAP, Director of the Swiss Center for Biometrics Research and Testing
24 March 2017 · 9:17 a.m.
Keynote - Reproducibility and Open Science @ EPFL
Pierre Vandergheynst, EPFL VP for Education
24 March 2017 · 9:20 a.m.
Q&A: Keynote - Reproducibility and Open Science @ EPFL
Pierre Vandergheynst, EPFL VP for Education
24 March 2017 · 9:54 a.m.
VISCERAL and Evaluation-as-a-Service
Henning Müller, prof. HES-SO Valais-Wallis (unité e-health)
24 March 2017 · 11:35 a.m.
Q&A - VISCERAL and Evaluation-as-a-Service
Henning Müller, prof. HES-SO Valais-Wallis (unité e-health)
24 March 2017 · 12:07 p.m.