Player is loading...

Embed

Copy embed code

Transcriptions

Note: this content has been automatically generated.
00:00:00
visitation all for texture that does it
00:00:03
so all the idea was to uh have um visualisation
00:00:08
of textual that asset and a possibility or for carrying a
00:00:12
such a a database is a totally visible
00:00:19
so usually you have this when you are performing
00:00:22
a search you know it very well uh ooh and
00:00:26
the output is kind of a linear information so
00:00:30
you um could have much mall but for many reasons
00:00:37
uh mostly money uh we uh you off this kind of output
00:00:44
so the idea was to uh say okay in an information space
00:00:49
we have more information like keywords that we can extract from publications
00:00:54
like also is that we can uh uh ah of organisation
00:00:58
and so on can we uh displays them in a way that people can interact with them
00:01:05
and uh uh having more uh having a
00:01:09
deeper insight on the uh that asked of course
00:01:14
i am doing it on a chance if you duplications database but uh we have uh
00:01:19
it's very vessel to in fact and we are uh doing it on a many different projects
00:01:26
in their mates even in a few that particle physics we
00:01:30
have uh some application so we have different projects that are uh
00:01:38
are that are on the run at the moment
00:01:42
so the idea is to beat each time a a coke you once
00:01:46
is that in the data set so co so school accurate and so on
00:01:52
uh that soul at the beginning where where where present things on on the bike
00:01:57
graphs and we have a um moved little by little the first two i paragraphs
00:02:03
uh which is just an extension of crafts uh to uh
00:02:07
a link between many of yes this is instead of two purchases
00:02:12
and uh now uh we are moving to what we call i. h. b. crafts all it's a um
00:02:18
in i perhaps you have a family of uh that are
00:02:23
composed of uh sets and now we have family of multi sets
00:02:27
so when we decide it's you can a duplicate elements but you can also put individual weights
00:02:34
and this uh allow you to do a lot of uh a lot more things in fact to complete the structure
00:02:42
of a hyper groups
00:02:48
so for instance and yeah the interest is out there and we can
00:02:53
we can do uh some diffusion on this kind of uh that asset
00:02:59
um on this kind of that worked the idea was to have a a
00:03:03
diffusion that is done in two phases and um one is still on the uh
00:03:10
that is is one step and user step is done on the h.
00:03:14
p. h. uh that links the the that this is a vocal currents
00:03:21
and so we age uh no i we had the journalists condom okay and uh
00:03:26
so we can wrong both that this is an urge th by this way and uh
00:03:34
it enables costing of h. b. craft behind and also a g. i. d.'s to uh uh
00:03:41
we find that the the output that we have in the text or way to all the it differently
00:03:49
so that's the next that it's not done already
00:03:55
so in a scientific publication the database you will have this kind
00:03:59
of data um you will have your articles a. b. c. d.
00:04:03
your organisation one side your keywords user side that you have extracted
00:04:08
or that you have the that have been put by the author
00:04:12
and so you are reading your coat currencies and you can visualise it and uh this what
00:04:19
here you are as a reference we choose the the publication itself
00:04:23
and we build the co occurrences uh of uh organisation
00:04:29
but we can do
00:04:32
something a bit different just putting ways
00:04:38
and it comes naturally when you take as referenced the roots
00:04:43
then you see that some organisation all curing many times
00:04:48
and you need to have individual weights uh that are touch low uh the um
00:04:54
to uh to the organisation and in this tax in this kind of a
00:04:59
thing and then you you need the h. b. graph two two two models and
00:05:09
then after we built a framework that almost to me but uh
00:05:16
to work into it i'm not i'm going in the in
00:05:20
the details but uh the idea is that we have the reference
00:05:25
and we have a physical reference which is usually the publication or any and they did this way
00:05:31
so there's a market in between the reference and the uh physical uh a reference and
00:05:38
we can buy this physical reference go
00:05:42
everywhere on every facet of the information space
00:05:46
and so from there you can vary what you want and
00:05:50
you can also change a preference and asking a different question
00:05:55
oh
00:05:58
you can retrieve your uh a draft of search and as
00:06:03
every node every vet x. it is uh uh interactive both uh
00:06:12
the the the vet uses of a of a lot of cocoa wants
00:06:16
but also the extra not that we put two symbolises the reference uh
00:06:24
we can see to your for instance you have fall fall fall um
00:06:30
uh that that this is not on link by the by a reference
00:06:35
and this references attached to the physical reference all too many physical reference
00:06:45
and so you can
00:06:48
uh explore your graph of search you can match graph of search of
00:06:52
different people you can save it and a replay it uh uh after future
00:06:59
so in fact use i i'm going to make a little demo uh we have
00:07:04
also put more information so people can uh uh uh
00:07:10
have additional information for also seen uh for instance you have a link to d. b. l. p. uh
00:07:17
for publication uh uh you have the abstract and the for an article on oxygen
00:07:23
uh uh for the keywords you can link them to uh we could video information
00:07:30
and uh uh some the the goal of uh does somebody creation uh one of the information
00:07:37
so all the cherries are made online and everything is processed online
00:07:43
okay
00:07:49
and incorrect is it's very vested because uh we can do it also just like having the uh uh
00:07:56
metrics of uh of me too sleazy to add so the incident matches uh of of the h. b. crap
00:08:03
and so from there we can uh change the reference varies
00:08:12
so this is just a little bowser of the the biggest one which is collaborations putting and from san
00:08:19
and um uh so the it's a project that is uh maybe now six seven years ago
00:08:27
uh it started six seven years ago and um uh it's totally it
00:08:32
this this one is totally generic and uh we're paid to difference to change
00:08:45
just like with the movie
00:08:51
uh_huh
00:08:54
yep
00:08:59
so
00:09:02
think of the last is this going to be hard for me
00:09:10
let's imagine that we are looking for something on a genus huh
00:09:16
and we want the fifty first uh has a and results
00:09:22
then you retrieved information so everything i we we uh carry oxygen so
00:09:29
you all the information that is retrieved from the l. c. v. p. i.
00:09:33
you have of course what i say the uh everything to gould to of uh uh the abstract
00:09:40
for two of the p. d. f. for the folks phones so it's uh this is not
00:09:48
this is common
00:09:51
but you can also interact so here you have audio so's users
00:09:57
are given by the uh uh sorry and maybe we should put it
00:10:02
uh_huh
00:10:04
whose can
00:10:09
and you skip church no i don't
00:10:27
so you have your your your view all sorts of course you can say now i'm interested i i
00:10:34
don't know in this collaboration so what you will see
00:10:37
everything on every fast set what is concerning this collaboration
00:10:42
and uh it will be a highlighted in uh in the different uh places
00:10:48
so you can see the keyword that where uh the most are represented in the in the publication
00:10:54
and also the subject category that is provided by uh oxygen
00:10:59
the curator not provided by alex if you have to process them
00:11:04
okay
00:11:06
then uh
00:11:10
so here you see you see that there's a a huge
00:11:13
part that is a very uh uh linked and there are some
00:11:18
oh clay that all players or some different subjects that are
00:11:22
yeah or not uh related to the what was gone for
00:11:27
so if i want to search on the gene and it's
00:11:31
uh it's 'cause of frail i just off to say okay
00:11:35
now i want june and schizophrenia so we could we compute automatically everything
00:11:43
then you will have your graphs and you can refine your search just what you can put or not whatever you want

Share this talk: 


Conference Program

Methods for Rule and Knowledge Extraction from Deep Neural Networks
Keynote speech: Prof. Pena Carlos Andrés, HEIG-VD
May 3, 2019 · 9:10 a.m.
1443 views
Q&A - Keynote speech: Prof. Pena Carlos Andrés
Keynote speech: Prof. Pena Carlos Andrés, HEIG-VD
May 3, 2019 · 10:08 a.m.
Visualizing and understanding raw speech modeling with convolutional neural networks
Hannah Muckenhirn, Idiap Research Institute
May 3, 2019 · 10:15 a.m.
367 views
Q&A - Hannah Muckenhirn
Hannah Muckenhirn, Idiap Research Institute
May 3, 2019 · 10:28 a.m.
237 views
Concept Measures to Explain Deep Learning Predictions in Medical Imaging
Mara Graziani, HES-SO Valais-Wallis
May 3, 2019 · 10:32 a.m.
455 views
What do neural network saliency maps encode?
Suraj Srinivas, Idiap Research Institute
May 3, 2019 · 10:53 a.m.
719 views
Transparency of rotation-equivariant CNNs via local geometric priors
Dr Vincent Andrearczyk, HES-SO Valais-Wallis
May 3, 2019 · 11:30 a.m.
457 views
Q&A - Dr Vincent Andrearczyk
Dr Vincent Andrearczyk, HES-SO Valais-Wallis
May 3, 2019 · 11:48 a.m.
103 views
Interpretable models of robot motion learned from few demonstrations
Dr Sylvain Calinon, Idiap Research Institute
May 3, 2019 · 11:50 a.m.
193 views
Q&A - Sylvain Calinon
Dr Sylvain Calinon, Idiap Research Institute
May 3, 2019 · 12:06 p.m.
125 views
The HyperBagGraph DataEdron: An Enriched Browsing Experience of Scientific Publication Databa
Xavier Ouvrard, University of Geneva / CERN
May 3, 2019 · 12:08 p.m.
104 views
Improving robustness to build more interpretable classifiers
Seyed Moosavi, Signal Processing Laboratory 4 (LTS4), EPFL
May 3, 2019 · 12:21 p.m.
132 views
Q&A - Seyed Moosavi
Seyed Moosavi, Signal Processing Laboratory 4 (LTS4), EPFL
May 3, 2019 · 12:34 p.m.

Recommended talks

Du Venise au Sion Time Machine : valorisation virtuelle du patrimoine historique
Frédéric Kaplan & Alain Dubois, Resp: Professeur EPFL et Directeur du Digital Humanities Lab & Archiviste cantonal de l’Etat du Valais
Oct. 2, 2019 · 11:03 a.m.
181 views
Why Google cannot be the # 1 in Korea?: In Search for Critical Success Factors from Local User Experience
Jinsoo Kim, Sungwon Beck, Sungeon Kim, and Kihyun Jung, KIPFA, Daumkakao, Naver Corporation, and SK Planet, Seoul, Korea
April 23, 2015 · 11:33 a.m.
135 views