Player is loading...

Embed

Embed code

Transcriptions

Note: this content has been automatically generated.
00:00:01
well
00:00:06
okay so let's start the last presentation maybe and hope for the lunch
00:00:11
so that today's topic is the interpretation of like the feature spaces for the want to meet in the combustion the network
00:00:17
and i using that by pressing on there for the part timers
00:00:19
things of the calls politically about which is used for the investor system
00:00:24
and these are the contents so first diary briefly introduced about the
00:00:27
what is the topic about and one of the many constant for this
00:00:31
application as a cannery just show you how to interpret
00:00:35
the one dimensional privacy than or lesser i briefly escapee
00:00:39
how to generate the payout were quite they access to the conductor thing and i've ever show you the research
00:00:45
so lester so the thing that i take your party you the for the application in the industry system
00:00:52
any deliver to rotate in power and the failure or currents and eat your costs all
00:00:57
curvy last obviously show an economy cost so the diagnoses for this matter is really important
00:01:03
in this propose are there lots of them what comes to adapt like physical domain work is commodity
00:01:09
for example there are lots of the attempts to using the
00:01:12
high frequency domain or something but it has not that good performance
00:01:17
in that rescind nowadays they are somewhat terms to use the motion morning for these
00:01:23
matter uh yeah there it tears you really good performance but there's no explanation why these are
00:01:30
these showed the really good performance for the by the most days
00:01:33
but the problem is that because it is see related with the decision
00:01:37
making rather replacing the components up the industry system or not as
00:01:42
a really high cost so explaining why this resources happen is really important
00:01:47
in this region i in this thirty i want to suppose
00:01:50
how to explain what the c. j.'s came from the network
00:01:54
and what is the result of that investigation and also for the television i use the two
00:01:59
different system up to date at the first noisy syntactically modelled as a common isn't really experimenter major
00:02:06
so uh let's see the basic idea is really simple cause i really using the one team is under signal or the vibration
00:02:13
cut and in that region hours us interpret this these them at the
00:02:17
entire neural nets up into the signal system analysis which means deprived because domain
00:02:22
suffers there's the commotion network i mean the compilation layer and d. as we
00:02:27
noted compilation layer can be interpreted as the a. t. i. f. fire fitter
00:02:32
which is just the to the the product autocorrelation cross correlation
00:02:37
and using that i think that the train the neural network can
00:02:41
be interpreted in the frequency domain which spend passes converse that we're not
00:02:46
a second there are some consternation about that when they are basically
00:02:49
for the image classification or something we use the mets play a lot
00:02:53
but for this matter we cannot use the max when you may be in polling because there are some random this
00:02:59
for the next building and mean falling there are some random is that which part of
00:03:04
the topic window is choosing is has a random and that and then is i think candle covered the
00:03:11
destruction of the signet essays on the singers so
00:03:15
in then resume i want to use the ovaries indicator
00:03:19
and over the pulling it says the composer with the two different part the movie lovers filter and the dismissive hated
00:03:24
and these two different parties it to the same procedure op
00:03:28
the until i think peter for the house simply and the distillation
00:03:33
and it is strong and here so maybe habit to different hyper parameter for
00:03:38
the movie over to the first one is try site as the commas colour side
00:03:42
and both are related with the uh last in band and the band test lesion
00:03:47
so we have to the allocated to decide what is the common size or
00:03:51
what is this trying side concerning the advice invent region and the convergence reason
00:03:58
and the last one is about the for the connecting the egg so for the for the for the quality
00:04:02
layer on the one dinner system the output up difficult to kill a year into the output mission like this
00:04:08
so basically here the ads which means that the final output op deconvolution that are
00:04:14
and the problem is the just the you know products are all the for the committee later i mean but it's
00:04:19
the y. is tied to the matter to the health status like
00:04:22
the old one rumour and this procedure can be expanded like this
00:04:27
so it yeah these are the kinds up is um changes using
00:04:31
the fury analysis where fourier series or something and basically doe final
00:04:37
because i'm on the finalise show that the fully connected it lay here can
00:04:40
be interpreted on the frequency domain again and they are looks like extracting stuff precursor
00:04:47
information on the specific weights weights
00:04:52
product i mean for example we can transform the
00:04:57
input data i mean the output op the convalescent network
00:05:01
like here into the frequency domain the largest uh we
00:05:05
waited a summation we use the weighty summation using the
00:05:08
wasteful before the committee lay here which is transforming due to t. free essays coefficient
00:05:14
and by someone some a summation op amp we can finally
00:05:19
circa what information easy safety from the for the committee layer
00:05:23
and they saw last i want to first show it b. p. r. called directories or is
00:05:28
simply think that it doesn't matter how that you how much the each comment on each in though
00:05:35
has an impact on making decision so we just the activate every corner and compared
00:05:41
the variation up the taiwan this is per performance and let's show you the data generation
00:05:50
these are the description about the how to be generated by person beat up the particular part
00:05:56
yeah it is set to make and everything are just prefer
00:05:58
skip and these id research topic generation and so basically though
00:06:04
generated to fire is a similar show like this so they're somewhat elation
00:06:08
and in the frequency domain they have some your big place which critic here mister frequency
00:06:14
yeah it's well known that that here is a frequency has a rape
00:06:19
oh i had the good indicator as they had they just for the cure politically about
00:06:25
so as you can see and hear there's three different present the first one is the the where cases a second laces
00:06:31
there's fold on the funk europe apply to get out and the third one is pulled on the politically about twenty here
00:06:37
and we can see that or and the g. m. f. area there are some fluctuation when the
00:06:42
all pro then the forties local and that looks station shape is very different hold on to his status
00:06:48
but the problem is in their your case there i disappeared house and visit and
00:06:52
they see their gaussian noise is just overall that this information on the free custom make
00:06:57
so we can at this thing is whether it is very it is no more
00:07:01
war forth with a real eye by comparing decide spends all around the general area
00:07:06
so this is samuel curse on the experimental data so these are the measurement top experimented a
00:07:14
recipe basically the vibrancy owner also should be clear high place on the g. m. f. area
00:07:19
well with the gaussian loses its just overlap so we cannot distinguish
00:07:24
so we hope that the our proposed to see and then we would discover the information from
00:07:30
the oval that'll easy to extract thing besides spend
00:07:33
information from that europe area there is our expectation
00:07:38
and these are the region that the result first it first of
00:07:42
all this is a a research topic a study about the pulling option
00:07:46
so as i mentioned before the collar side and the strain number and
00:07:50
the other is pulling beside the convergence area and the anti r. i. scenario
00:07:55
so as you can see in here if we increase the collar side then the overlapping
00:07:59
area i mean the converters erase become there were and this show the clear narrow lopez feature
00:08:06
and also if we increase this try size been the allies
00:08:09
in bandy really increase the beach candle clarity of icing problem
00:08:14
and as we can show in here about the performance if really is about other qualified
00:08:19
astride number for example if it is the common size eight reaches the yellow line here
00:08:25
and just try size to which is done till they spend here alas invented here
00:08:30
they and there is no interruption between allies in because the low pass filter is
00:08:34
enough narrower so it can be at s. b. and p. l. s. a. filter
00:08:38
and as a reserve they really show would be the height performance and convolutional network
00:08:44
and also after that we just compared the performance about that i mention is the key are
00:08:53
all along to the corner so there are three highs the purer corners and the last corner
00:08:59
and we can see that the decrease that raises a really very different
00:09:03
so the highest g. p. r. connors impact highly on the making decision and
00:09:11
if early
00:09:14
if we investigate the free can share free concert
00:09:20
uh in the frequency domain or condor then you can see that
00:09:26
the collar flu showed beehive sensitivity i'm you hide p. r. r.
00:09:30
show the clear condor dish or on the band pass filter around each enough area
00:09:35
bert uh but that the kinetic which showed the voiced c. p. r. r. b. f. data conversion see
00:09:42
so as on here we can say that okay then if we train to c. n. n. l.
00:09:48
and they will convert some of them converts into the area around the chimera
00:09:53
and so on our tail to convert it and they are highly impacts on the
00:09:57
making decision so these are the results about the on last season the compilation that yeah
00:10:04
and finally this is about the for the committee layer if i mention before all the way so before the
00:10:09
committee layer can transformed into the frequency domain again and
00:10:13
as we can see and hear these three different uh huh
00:10:17
the way so long to the highest here are colours as
00:10:22
say a little study ways along to the voiced p. r. canal
00:10:27
i'm sure they sure really senior shape on the frequency domain although the pair are so different
00:10:33
so from here we can just the shrink the hour and maybe kansas
00:10:38
estimate our previous occasion into here so from now we don't need to consider about the
00:10:44
the conrail where they're they showed the high high p. r. or loci are
00:10:50
and the result is as follow so overdone rumour cases and doesn't give full employee i think you'll fall
00:10:57
they should be clear saying formation on the frequency domain
00:11:00
along to the fulton will i mean the classifier what
00:11:04
and is equally saver do real time and the engine or tong all the way
00:11:09
so from here we can conclude that the full equality layer
00:11:14
in the frequent for the collectively here to always try to extract the feature information and we keep feature
00:11:21
from the frequency domain along to the baltimore and days show the different shape along to the fall to more
00:11:28
so these are the conclusion of the for the quick earlier and this thing fundamental yes the cost that is
00:11:34
the investigative eon be experimental data so for the these
00:11:40
are the result of the comparison all the p. r.
00:11:43
and these are the rules thing or on the vibration so
00:11:47
these are the no more and this sunday afford at twenty fourth
00:11:52
and the issue to cornell shape on the frequency domain and here as you can
00:11:57
see the canal to show the highs p. r. r. article the converted to degenerate area
00:12:03
but dope connor to show the last p. r. r. also fail to be
00:12:07
converted into d. g. m. f. area in full art as a previous action the
00:12:15
the output stop i mean the frequency responses stop the for the committee layer
00:12:20
also you call the shape whether readable they stop e. p. r. i. so these are
00:12:25
the most as a pure uh this uh the list of lists is the v. p. right
00:12:29
and they show the same frequency response along to the
00:12:33
classifier mold on the near term and the user tumbles
00:12:37
so we can conclude that okay so the first the pulling here
00:12:42
but if we use the order is probably pretty one division nurses them then we can consider the hyper parameter
00:12:48
onto the commerce size and the stress that beaches along to the convergence area in the hour until i sing band
00:12:55
and second the convolution layer and interpretive it in the frequency domain
00:13:00
again maybe shoot that the corner which you should be high stick
00:13:04
p. r. conversion in teaching the fairy exactly but the connor to show the listed
00:13:10
p. r. is fair to convert g. e. or around each interfere yeah
00:13:14
and so are the response of the product conductive layer is converted into d.
00:13:19
the feature extractor along to the same feature remote about along to be
00:13:25
we classify remote and yeah if worse as the the
00:13:30
distinguish the sink to the east feature extractor on the classifier
00:13:36
so the conclusion is that related features basis to you this in a using to to really interpretive using
00:13:42
the constable tuner system on this is and they are equal to the peace corps and all these that
00:13:48
that she information is converted origin if area and a commercially are are sort say
00:13:54
it to converge into the g. m. f. area if the tissue the highest p. r.
00:13:59
and if they shoot a list jar and then they they to convert your and h. m. f. area
00:14:03
and also the past perfectly layer is acting as the
00:14:07
feature extractor we chase different along t. d. s. classifier remote
00:14:12
and yeah the process of or is that they the using the fight the we're cases i mean these
00:14:18
payments are cases and this is the case and balls are show the clearly say reserved for the can present
00:14:26
yeah i think this is the record my presentation and cause i

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.