Player is loading...

Embed

Embed code

Transcriptions

Note: this content has been automatically generated.
00:00:00
alright and see my can you know i'm from the whole but learning on interaction group
00:00:04
so but originated from large scale manufacturing the
00:00:08
day on now widespread to wider range of applications
00:00:13
he manufactures we have about that are much closer to the to
00:00:16
the users so well sharing actually the same workspace as the users
00:00:22
but also many more applications including all boats that are
00:00:27
assisting us but also supporting us about being part of full body
00:00:32
and also but being far away from us that stupid operated by people
00:00:37
the in between all of these applications is that we
00:00:40
require to report on these robots much more frequently than before
00:00:46
so it means that we have this whole boats that need to adapt to new tasks
00:00:49
to use else to nonviolence and for this reason we are interested in
00:00:54
learning from demonstration mechanism has a way to transfer skills to go about easy
00:01:00
so there are many underlying challenges first there are many ways of uh acquiring date are
00:01:06
from the demonstrations is or from views your observation of form kinesthetic teaching
00:01:13
we also menu generalisation problems how to adapt to the task to new
00:01:18
objects to new situations to move about two new persons of to new environments
00:01:25
so this is the typical example of experiments but we're conducting in our group
00:01:29
so we teach new schemes to hobart by guiding van with the difference
00:01:33
that of the task why the whole body's recording various kinds of information
00:01:39
and we often do via an interactive my no in the sense that we are
00:01:44
providing demonstrations we'll observing the result we
00:01:47
are horrifying the demonstration on sometimes only partially
00:01:51
on we have this interaction process where we can also mm visualise what the whole but as well
00:01:59
so i still does kind of the ministrations this is what we what we what
00:02:03
we have after we have about that kind gentle eyes the task to really shallow
00:02:08
of the post rules we have about but internalise the task to new target position of the object
00:02:14
and we have about the country lies a task to new a composition of these obstacles
00:02:21
to my presentation i will um why we show that the the main
00:02:25
challenges but we're facing is that we don't really have database of these movements
00:02:30
so we don't run on screen the set type where we have like a like like testing uh like train
00:02:36
beta testing data it's more but we uh can acquire
00:02:40
the date are uh by uh providing smart interaction mechanisms
00:02:46
so yeah we we need to start learning from the beginning soon to
00:02:49
start running only when the few demonstrations are white able so this is
00:02:53
the meeting the number of techniques you can use this is also put
00:02:56
a strong emphasis on the model priority but we need to to us
00:03:01
so we would like piles but are still valid for
00:03:04
a wide range of cost without knowing the task in advance
00:03:08
uh but we also need that it's not the whole but
00:03:11
start doing something only uh when should instructions or white label
00:03:16
so the interpretation capability is yeah at several levels
00:03:20
so first at the level of the interaction process
00:03:23
as i mentioned also the level of these trials so i will i will start with the interaction perspective
00:03:31
so yeah we first have local to t. in machine learning a little bit pieces that we
00:03:37
can also consider several forms of real learning
00:03:40
on several forms of mechanisms we're actively only mechanisms
00:03:44
we also have machine did she where we can tackle the problem
00:03:49
of generating data that would be the best to train all models
00:03:53
we have to recruit on learning where we are interested in the sequence at which we
00:03:58
put his data on there so how does the off targeting dialogue now
00:04:04
and more generally we have some interaction mechanism so oh to explore its
00:04:08
social uh interaction capability to have bet a husky tones fills the mechanisms
00:04:15
to this boils down to to uh to to trying to interpret modelled by
00:04:20
interacting with the whole boat and i like to see that as a scaffolding process
00:04:25
so it's like a metaphor to show but at the beginning we need a lot of structures
00:04:30
on once the whole buttons acquiring the task we can remove progressively these structures
00:04:36
so if you want it's like the continuous evolution from for assistance to fullerton on it
00:04:43
we can also play with different learning modalities the whole but can learn from different users
00:04:49
they are different imitation mechanisms we can try to copy the actions that we can
00:04:54
also try to reflect about these actions to understand what was the goal behind these
00:04:59
on the knitting doubleclick included bores instead of the specific set of functions
00:05:04
we can also have about trying on its own so we have these different learning mechanisms that need to be orchestrated
00:05:13
no from or a technique al perspective i would say not the lever of
00:05:17
the models we we have several perspective that we can consider first when we
00:05:23
ah talking about movements what we are usually considering
00:05:27
is times every is located two different coordinate systems
00:05:31
this is some kind of prior knowledge but we can put you know
00:05:34
says than on but will help us at making these models into the table
00:05:40
so we start with this aspect
00:05:42
now we'll industry that was an example in the project called either has
00:05:47
well the goal was to have about helping people about putting guns shoes
00:05:52
so yeah what it means with this kind of scenario is but we need to consider the coordinate system of the whole bought
00:05:58
but then the movement need to be to get a two well the shoe so we need to consider also the coordinate system issue
00:06:06
and the coordinate system of the user so we are these different
00:06:10
kind of up several that we participate in the understanding of the task
00:06:15
so in some way or on the systems are doing some statistical analyses but from their own perspective
00:06:21
is it is as if you would put one up salvo on each of these got it says that
00:06:26
but you would provide some arguments restore its demonstration of the same thoughts on you would
00:06:31
add these op silva of reflecting about what the task is about uh from day one perspective
00:06:38
so it means they will up so valuations but evaluations will be different depending on the op so well that you would consider
00:06:44
two in this example we see in red the what the whole boat would like the movement to be
00:06:50
in room what's the shoe would like the movement to be on yellow the fusion of the two information
00:07:01
so we're exporting button orders project yeah i would like to emphasise one cortex hot
00:07:07
where we did operated the whole button and the what the whole bought these two arms from the data operations and uh
00:07:14
and the problem with to face here is but we had multiple calling it
00:07:17
says stands but bo from that their operations side on them before but site
00:07:23
and it meant that we had to face this discrepancy between the situations uh on the two sides
00:07:30
i knew this mechanism the uh by having the movement with
00:07:34
the multiple courted systems was taking us to cope with the discrepancy
00:07:39
yeah and and and that may shown you see how
00:07:41
the model is locally adapting to change of this coordinate system
00:07:51
then i would like to discuss or some geometrical perspective next
00:07:56
we have the chance in her butt cheeks twelve date are lying on well known many ford
00:08:02
it's the case fall the joint angle trajectories of a whole boats but
00:08:06
will be articulated drawings on but that uh some way moon geometry is
00:08:12
it's the case of the end effect also the end of
00:08:16
the whole bought that would be located in three d. space
00:08:20
is the case for orientation data or more generally rigid body movements
00:08:25
but yeah also as a a way known geometry that we can exploit
00:08:31
little also also maybe a mean less waiting room mechanisms articles like sprite
00:08:38
example saw many pretty change so it's induction mattresses stiffness mattresses
00:08:44
that we can use it for we have call volumes features
00:08:48
on all of this is represented as cynically positive definite matrices on
00:08:52
yet again we have the way noon geometry that we can exploit
00:08:56
so when we collect data are we try to keep
00:08:59
this geometry uh i into account when we are learning something
00:09:04
so it's valid for a statistical analyses so if we would like to do some clustering
00:09:10
well if you would like to infuse information from difference installs repeats it also the case for people
00:09:17
so we exploiting border situations and yeah if time only to show
00:09:21
one example so i picked up this example from the tact and project
00:09:25
which is about um and i think electronic often they taught to conclude prosthetic and
00:09:32
and in this application on what we re doing is to pick a sliding
00:09:37
window like a time window what was move the where we where computing special currencies
00:09:45
and when we do that a little more hold just way of treating
00:09:49
these data but it means that we had at each time step of koreans
00:09:54
on your we wear a developing location to worse but could take
00:09:57
into account the shape of the money for to process his data
00:10:05
not very related to the to the geometric aspect here is the structural aspect
00:10:12
what i mean here is that we have date are organised as multidimensional arrays
00:10:18
so full but the same can be or rate of sense all it can be multiple channels you
00:10:23
can be simply the time evolution of the signal you can also be multiple coordinate systems as i mentioned
00:10:31
so this uh no that's to look at the feet of
00:10:33
ten so much thought it's also sometimes called merging larger brought to
00:10:38
keep the the outside the simplicity of the techniques but we are
00:10:42
using because we need to have to to learn from few data
00:10:46
but but also exploiting the the the the mean the multidimensional aspect of the data
00:10:53
so if we if we would not use this what we would be tempted to
00:10:56
do was to take these multidimensional or rate on pause from but in effect all
00:11:02
which we would do but just this would be the the the typical kind of regulation program but we would have
00:11:07
uh we some input x. we some weight vectors w. some bias be on some output y.
00:11:15
uh in this setting we would always of the joint action program or
00:11:20
logistic regression problem depending on what we would put as the uh the function
00:11:26
so now let's see what we can do if we try to
00:11:29
extend that to morrow structures so if we have matt thinks they tell
00:11:33
you we can still keep something very similar are but this time we would
00:11:37
have weight vector all that can be applied on the two side of the mattresses
00:11:41
so it's like during the matters much breaks on one of the side or the other
00:11:47
uh and we can her face but as when first note of product but we do with
00:11:54
the weight vector halls on then in our product that we do to compare these different pens pencils
00:12:02
and if we take this whole presentation it means that we can actually
00:12:05
extended to tense all presentation so to control i uh huh the dimensions
00:12:11
so yeah this is the or presentation of what it would mean so can view but uh some
00:12:16
form of decomposition was the stone article position that we know for matrices but applied to ten cells
00:12:23
and it is interesting here is that we in the simple case we have fans saw hank one
00:12:30
but we can also extend these two uh should deposition of ten stores offering one
00:12:36
and what is interesting for us with this kind of a presentation is
00:12:39
not we can be in a complex of uptown b. is on the subset
00:12:45
off the easiest back down on we can highlight something to hide all of these in the vineyard but don't sit separately
00:12:54
it's also important for us because we can learn with very few demonstrations is these compared to work
00:13:00
vector hydration or for complete d'etat it's much less barnett else to be trained
00:13:07
so we use that as well in bars project you know i'm also
00:13:11
presenting an example was attacked and project it's his somewhat but with the the
00:13:16
so that i can project is about uh i'm uh i'm
00:13:18
controlling the prosthetic and for a fall i'm uh i'm computing is
00:13:24
but we also test but interpretation since sitting uh uh with a whisper some saving
00:13:30
faith they they and to your the idea is to look for this and that
00:13:37
the signals coming from a wood tactile graphic data so we have a bracelet reason the
00:13:45
ray of sun stalls collecting does a date are and what we need is to explore it
00:13:53
but if you know like they're shown on the logistic regression to be the the mixture of expect model
00:14:01
but the mixture of expert model that is taking into account the
00:14:04
towns all the top who've for the gate on for the expert
00:14:09
and we could sure but with this kind of a presentation we could have
00:14:13
both interpret table models on we could also train of system is very few demonstrations
00:14:23
so that they can win message uh what i would like to convey
00:14:27
is that uh i know but takes we can use ever are learning
00:14:31
compatibility on teaching capability it's interactive on we should take
00:14:35
into account thought when we design a or whatever reasons
00:14:40
it's also very important with the whole boats being transparent because since it's
00:14:44
interactive it means that the user needs to know what the whole button nose
00:14:49
so the reason is to assess what is the current level
00:14:51
of understanding of the task in order to provide that that'll demonstrations
00:14:57
then it's important for movement to consider multiple coding systems
00:15:03
in order to adapt quickly to situations and new users
00:15:08
and it's also important to take into account the structure of them there's only trees on we have the
00:15:12
chance to have quite clear data structures and quite
00:15:15
clear geometries but we can exploit meaning but we can
00:15:19
craft learning algorithms that can learn from few interactions
00:15:24
so yeah i emphasise mostly the long form demonstration aspects but it also
00:15:28
valued a wheeze learning by exploration was dilbert learning on its own don't aren't
00:15:35
we'll put the most awful source code online this is the learning a new direction group when you can see more applications on my website

Share this talk: 


Conference Program

Methods for Rule and Knowledge Extraction from Deep Neural Networks
Keynote speech: Prof. Pena Carlos Andrés, HEIG-VD
3 May 2019 · 9:10 a.m.
349 views
Q&A - Keynote speech: Prof. Pena Carlos Andrés
Keynote speech: Prof. Pena Carlos Andrés, HEIG-VD
3 May 2019 · 10:08 a.m.
Visualizing and understanding raw speech modeling with convolutional neural networks
Hannah Muckenhirn, Idiap Research Institute
3 May 2019 · 10:15 a.m.
Q&A - Hannah Muckenhirn
Hannah Muckenhirn, Idiap Research Institute
3 May 2019 · 10:28 a.m.
Concept Measures to Explain Deep Learning Predictions in Medical Imaging
Mara Graziani, HES-SO Valais-Wallis
3 May 2019 · 10:32 a.m.
What do neural network saliency maps encode?
Suraj Srinivas, Idiap Research Institute
3 May 2019 · 10:53 a.m.
Transparency of rotation-equivariant CNNs via local geometric priors
Dr Vincent Andrearczyk, HES-SO Valais-Wallis
3 May 2019 · 11:30 a.m.
Q&A - Dr Vincent Andrearczyk
Dr Vincent Andrearczyk, HES-SO Valais-Wallis
3 May 2019 · 11:48 a.m.
Interpretable models of robot motion learned from few demonstrations
Dr Sylvain Calinon, Idiap Research Institute
3 May 2019 · 11:50 a.m.
Q&A - Sylvain Calinon
Dr Sylvain Calinon, Idiap Research Institute
3 May 2019 · 12:06 p.m.
The HyperBagGraph DataEdron: An Enriched Browsing Experience of Scientific Publication Databa
Xavier Ouvrard, University of Geneva / CERN
3 May 2019 · 12:08 p.m.
Improving robustness to build more interpretable classifiers
Seyed Moosavi, Signal Processing Laboratory 4 (LTS4), EPFL
3 May 2019 · 12:21 p.m.
Q&A - Seyed Moosavi
Seyed Moosavi, Signal Processing Laboratory 4 (LTS4), EPFL
3 May 2019 · 12:34 p.m.