Player is loading...
SGAN: An Alternative Training of Generative Adversarial Networks
Tatjana Chavdarova, Idiap Research Institute
Thursday, April 19, 2018 · 10:32 a.m. · 09m 07s
Generative Adversarial Networks (GANs) represent an impressively powerful generative model, which is based on deep learning. The quality of the samples produced by this algorithm, made it applied in wide range of computer vision problems. In spite of this success, GANs gained a reputation for being notoriously difficult to train. We consider an alternative training procedure, named SGAN, where the final pair of networks is pitched against an ensemble of adversarial networks, whose statistical independence is carefully maintained. Such an approach aims at increasing the chances of a successful unsupervised training and improving the performances of the produced generator, in terms of coverage of the targeted distribution by the modeled one. The experimental evaluation also indicates improved stability throughout convergence and faster convergence rate.
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:00
ah i'm i'm going to present i'm not sure it's adjourned recognise the provider it's cool it's gone and it's i just
00:00:07
uh in a metal you or a different uh way of training out there that they share it
00:00:13
so before going on in here i can i just very very with the things that i guess but
00:00:18
i don't think so much about that and get out of it but i got it right
00:00:22
so basically it is the only way our goal is to train and given to model
00:00:28
we now what you read well which is all the discovery
00:00:31
or discriminated retraining our bases are simple on the generator
00:00:37
and on the real data sample from the real data distribution would you want to learn
00:00:43
so basically we have implicitly bigger labels assigned one coming from
00:00:47
the real beta words you are coming from the generator
00:00:52
and then the train the generator that generates a in a link that this committee that the simple sound like real so
00:00:58
basically it's strange because then i'll make the label because it level lock ups so because we're always this like this
00:01:07
ah like it's really interesting is that the generator like in the out and go there
00:01:11
uh we thought it's a mapping from what they expected to the real data space
00:01:16
so these are not expect their example on some disruption within which we can easily sample
00:01:22
so we can form like that like this is he saying that because there is a shot
00:01:27
and i think really bad we know that the temples uh this is just an expression more than once the
00:01:33
uh as you are some because we have the players one of them is trying to
00:01:37
maximise the value function another one is trying to minimise the that function so basically
00:01:43
uh i think maybe we're of in thing you really didn't show
00:01:48
there are a lot model this division is going to be able to relate the story
00:01:53
so right but uh you can interrupt the necessary so this is just to illustrate your load but may
00:02:00
just uh basically this coming either it get examples from the related distribution we get in the labels
00:02:07
examples from the generated the generator is getting variance from the discovery so that's how to generate a model is thing
00:02:15
okay so uh this is when you generate images and balance level they found so many of digits i'm not going
00:02:23
to do with it i just thought struck the size wary focusing on that but it's pretty clear uh_huh
00:02:29
by the number thing but that had some big i mean that that boxing and does it grows exponentially
00:02:35
so the the tray ah although it's very popular topic which are useful
00:02:40
oh in many applications it's there's so many problems so there isn't
00:02:45
that your top labour you really want to use that
00:02:48
um so this is what we of b. b. c. it's happening but it's a lot more complicated basically
00:02:55
the parameters of the discriminated the image and likeness and uh like one one
00:03:00
parameter just working out the main dish and what you wanted to
00:03:04
then when your devised a valid points that they shouldn't be doing this line we actually once you find them that all points
00:03:11
which is going to give us the optimum the optimum else they i think
00:03:15
that one but this never happens because at varying on on its use
00:03:20
um and there's the neural network of bearing on what's in chains there's so many so many problems one
00:03:26
of them or menu then what you want one mentioned that your slides i will just you
00:03:32
all this and like this is samples examples of what collapse and it's not so basically
00:03:40
uh they have been so many um variance of the garden going out with them but it's very difficult to
00:03:47
say for that one and then outperforms all the rest although they have been doing someone's someone's work
00:03:53
because they're they're a little higher priced and oh and depending uh said get get get different performances
00:04:00
so what if i guess it uh is sewn on different mythology how we can uh can you hear me or oh
00:04:08
so the parental as you probably can train a login other side of the brakes
00:04:13
um and this can be applied to any variant of
00:04:16
that cannot within that already exist so oh
00:04:21
okay so it is motivated by you know some
00:04:24
empirical observations so basically if impinges so
00:04:28
uh run multiple insinuations of of of of the father stereo pairs
00:04:35
we will see beneath the convergence of around the mote so you are real i'm not sure
00:04:41
i can see but if you are a rerun experiment where we just they um
00:04:48
oh very train one it's the one pair of other senators regularly
00:04:53
and we train another one but the this one the second
00:04:56
one estrangement stumble off well and symbol of well
00:05:00
such adversarial paris we we didn't want to one that does give me the screen their output
00:05:07
in some region of interest this is what we get the sit there
00:05:10
for that generate their card it's up for this game either
00:05:15
a bit of a cripple just come in there that is some uh change regularly there's what
00:05:21
collapse and tons of the way that we train a a there there's no what colour
00:05:26
uh the probability that a problem or it will it will not be hard right
00:05:31
at the number of independent adversarial problem out here to our
00:05:35
guest goes down exponentially somehow can use this uh
00:05:40
so this is how the structure of the algorithm it's basically that have shaped our job with them in the end
00:05:46
but there is one pair networks and uh oh which is very convenient for practical
00:05:51
applications although you use many other smile appears again you then just one pair
00:05:57
but here's our insensible so basically the is trained to discriminate older older
00:06:02
examples and the generator is trying to pull all the display niggers
00:06:07
domain we as tuned to the the diesel here's the tape all
00:06:12
uh in our uh our main thing is that when when there's
00:06:16
no one network things lessons them all because we didn't get
00:06:20
what's collapsed or steam out okay so oh basically this is
00:06:24
the essence training and can generate than just one pair
00:06:28
so and it not only improves what were surprised that it not only improves our remote coverage
00:06:34
but it all seemed plus converges sun rises type and you say the blood samples from
00:06:39
a faint from the length examples here this is the real data distribution
00:06:44
and this or samples all taken from the different local oh local uh generators
00:06:50
indeed and they're not that in different colours so that little generator
00:06:54
convert just in a different way than the local ones
00:06:58
and they just so uh see i like that the next generations
00:07:02
we see that it's like seventeen the the local and so that the trajectory of converging is different
00:07:08
the best therefore the faster convergence it's is really interesting okay
00:07:14
um hum some metrics when you get like a different brands like that we've done of this again
00:07:20
um dark brown over other buttons on any brian you mean then when the the performance is better
00:07:28
so i want to show some also it helps what collapse and uh though and i mean what we see
00:07:35
on copper experiments it also happening and you know the the samples what this very interesting is that are
00:07:42
just selections are also reduced and i'm i'm sure just few years
00:07:50
happen
00:07:52
so i asked the president bush does that big mama generator and this is the start of local ones
00:07:59
so basically they're all of us like teen way more that i'm not very conceited probably the resemblance
00:08:05
but there was telling you can definitely see that there ought selecting a quite a lot like about
00:08:10
so uh in the blue r. square that's that will generate another you train and the other ones are regular actually
00:08:18
and uh basically they are sometimes does oh like you can do this for multiple of data set
00:08:25
and you say that you're due to that selection sometimes or the local paris it's going to it's going
00:08:30
to completely fail the um the convergence various target level and we always get this and we also
00:08:39
you mean
00:08:41
even his show it i mean it's not just some parts of their comments
00:08:45
to because the scene but the convergence is a lot more stable
00:08:48
so basically this is what we get and this is our agenda that has been
00:08:53
this is just to equalise computation although although although this
00:08:57
there's more defined benefits so just you just you
00:09:03
yeah
Recommended talks
Deep Supervised Learning of Representations
Yoshua Bengio, University of Montreal, Canada
July 4, 2016 · 2:01 p.m.
2369 views
Brain MRI Analysis and Machine Learning for Diagnosis of Neurodegeneration
Dr. Ahmed Abdulkadir, University of Pennsylvania
Feb. 6, 2020 · 10:04 a.m.
163 views