[00:00:05]
>> Ok. I everyone thanks for joining us start with a quick introduction today will be hearing from Professor bullies who is currently an assistant professor at the Chinese University of Hong Kong prior to that well it was a Ph d. student with Professor Antonio Taleb his dissertation work was titled interpret it will representational learning for visual intelligence many of you are probably familiar with his work on class activation maps or cams for short or some of the more recent work on work that's section both of which provided us with a better understanding of what was happening inside of our c.n.n..

[00:00:47]
Poll is won numerous awards and fellowships including mit tech reviews top innovators under $35.00 in Asia Pacific region and the Facebook page deep fellowship today Foley will be talking about interpretive will in space and the 1st problem in deep generative model. Is with. Great yeah thanks for that introduction.

[00:01:11]
Hi everyone Phil very nice to have this opportunity to meet you all you guys you know there's been this. Quarantine situations he's good to come back and go with people and make new friends so today how with it I have to talk about the series to work my group that I working on the interpretation all deep generate the model I want to talk about in terms of the bone that has faith and value versus call that are against.

[00:01:44]
So you know over the years we see great progress in the future with all the musicians assigned to 1st in 2014 again came out now over the years to receive the image of quantity and come ton of the generated image have been greatly improved over the years so we can see we have the style again the p. Ken style can be true that comes to emphasize all kinds of different all to ask him or hi fi that it's so even people cannot differentiate that rather it synthesize the image all the really image and also there is this very reason to program the images into science the 1st this time there are new role value rendering that can.

[00:02:32]
Like nurse coming criticism your reprint patients on the Thames do this. To generate some different views on the fly also this month we see this open the i Pad 2 units thoughts that can turn Anding recent synthesis into the game each to conceive here we have a. Chair in the shape of about how the Congo has been quite amazing that if we don't have to work can add their own creativity in creating a new image to have such amazing the present reading models go we want to understand why it has to decide such a release of the image so it's a relevant to the topic of the day that here we have a deep gender or patient so you know previous to open the eyes bullets they have read in school parent years although he says self giving nods or having larger though it can seem to Sizemore one more realistic needs to read try to understand why is this really just that you need to come to synthesize and also giving this stuff reprinted some of the presentation the what's happening there are the insights so here I show of very basic and generators so here we have again generator for this generator basically he's just a commission on the own network so we have some improved need him back to Earth so that you could factor your yearly sample from some random distribution so the stuff I call the random factors of the inputs then throwing to the general returns on the generator can see the size of the reduce the image because here we have a generator that generates paths through those out to kite so they have to rip and patients in this case the 1st type so blatant rip and has some awesome 800 units so here I call this combination of futures in the generator as amazing units there are not that are made from space.

[00:04:32]
He sees this need some doctors some pulled from this random distribution facility out of the Nathan space toto wards the standing and be sure lighting up the Nathan units there is this some amazing work done by Davey ball and on point with rock off from my previous lap that's to try to visualize the human internal futures through the show that's when we train in this generator to synthesize the things that can turn your need to actually merge adds.

[00:05:03]
All Jacked synthesizers So some union specialized who seems to side the clouds come to seem to size of the hole in the grass so that we can turn on the turn all the units activation then we can control this generation process to on the right I'm sure you can track the interface that we've come to her on the troll the units with the semantic meanings down adding on removing the come to.

[00:05:29]
Thoughts Hala boats to the space you know this place in space far out try to alternate in space. Actually has a tree in them for a folded generation process we 1st a sample of the random actors from the start you should be shown on those images the includes one through this generator that is consumed as I release the image than the poverty is what have been encoded in the bird beginning they sense the whole country I think defies the stuff they face in size this is a time space so here I have created 1st have a very unique show experiments in the studies in space but here is just the random walk in space to basically each time we just add some increment on top of the Navan back to us then you can see is the author who the me is actually changed according to this random program per permutations of the big game so you can see the same size the images come from north of different semantics they have different views and all through the number of bad will change and the texture of the sheets and the even the passed around watching Joe quoting 2.

[00:06:43]
Percent for mutation are in the state and specs so that makes out sync the slaves in place facing the u.s. in coach or kite hold on semantic the concept of cement you commission about the. Rules you pretend to back then defies us meaning for subspace that we can actually hear or control this a generation process.

[00:07:06]
So you know it's giving this nature seems a. Lot of semantics inside the major cities so my previous work on my kid's missis's work so I work on this visual scene on the standing so for this reason see on the standing is very caustic and hope you can computer vision them people feel it's different.

[00:07:28]
Most extractors who get all the mean the phone semantics call and this includes the image for example for what we have the city each novel enables So this is up as a rule and we also have difference I trip you the major likening with ties and also we can generate this.

[00:07:48]
3 d. a structure all you can see is 2.5 dimensional structure you call made from the layouts to describe the things we also can apply the semantics mentation of treat shots the all objects and their regions so you can see for this team facade an image all kinds of different semantics so you recounting those semantics meanings back into the Nathan space that we can control this generation process also a number of very interesting things out so when we trim the small code to do this kind of beam to synthesize we don't have any ob there are external supervision the system convert to synthesize the image by itself to whether this interpretation or reading nerve or those kind of the different multiple novel obscene descriptors is also on the long questions.

[00:08:39]
So along with my students young and the Shun So we propose some crown work to identify the causalities in the data space so basically saving this generation pipeline we come up by the off the shelf predictors to extract many a semantic informations from the things the sides imaged and then we can go back into the nascent space and.

[00:09:06]
Then defies the course you factor relations you know you recount them to fight the dimensions of the subspace and then they can state that can control the relevance of the final goods thing and we can control that generation process so we propose the frown work on the very simple from work on Tweeter preach the nation space to the city we 1st generates notably some vectors from this debate in space by sampling from the random.

[00:09:36]
Distribution So just a long post on distribution and then we can stroll into the generator then this generator generates a lot of images that we have this even the space that we all fly some off the shelf clock to farce you know from the place come or not it's previous that Felix or some of our songs sing attributes cut the car down we can predict for now tributes for the states aside the image to hear reaches the top some in the closet are that we can have the label cold weather and so even though all life and then all the whole life in competitions so we treated this predict to the label goes to the label for the lead him back to us now we can go back into the listen space Duntroon some classic cars 3 deter outs this mean you're on to the concept our current chief a pretty life Cousin occasion a Creasy for this fine or a classic issue you know whether it's the end all although that was how it should have been more than 90 percent not p.c. by using the pseudo labels there is also shows out this nation states have laws on the scene can go the semantics the already existing time go to France and I think therefore we can achieve such high accuracy so we have another very important stocks go up here I call you tell coming for fracture verification it means we have to make changes I made some code timed up with over the output.

[00:11:06]
Riser this boundary really control this to us to have this of replication step it's a very important step after we add to some increment still this boundary and to see rather this image after bills have been changed now 10 years just to measure how well these control on some semantics those through this process we can then define the reliable Congress in the nation states countries with different semantic meanings to here we 1st have a sundries that country control is a major lightning and they know lightning so here I would like to show you some mission as a nation so let's see if we have 2 random vector of a time space that's beginning they have.

[00:11:52]
To look like this you know you need to arm the image you can see there is the lamp use ice so here we gradually prove that made him coats unter was the boundary still by adding just of the normal of this. So you can see the life in condition and gradually change the land on your side the scenes of her down the.

[00:12:14]
Series the more that cause damage video all this to you can see a very beautiful this continuous change all of the sudden lighting conditions the lows on the ng the lamp is turned on also the globalizing condition you go searching so you can clearly see the shadows projected on the wall so it's much better than the simple and Photoshop results in this made in space more coming we have 25 though I think we can all start them defy many odds or attributes for example we can define the boundaries for clouds and the long cloud and the boundary school education and the longer decision than on to read them to follow those different bung the stone we can change their semantics to here I'll show you another time a video of aiding cults going on here we have some images so it's have very dramatic backgrounds so the rather condition is not that well right now we want to have some clothing insight.

[00:13:21]
Who carried the video that we just proved the latent voted for was a cloud boundaries so you can see those clouds are more like how to not get the girl out in the sky for these early for the 2nd image so I really like it because you can clearly see this very beautiful sunset cover projected on the clouds as well as on the beauties there is a very realistic a manipulation so here is some other results that we have some buildings.

[00:13:52]
It's very boring because it's not the education that we want ads to be just his insights that we can push that it Nathan codes towards features ancient boundaries then the cross and the trees I just more like the mask of a girl out in the city also very interesting you can see for the 1st example so you can see this that generates very large enough to put the sidewalk on the grass because there is a sidewalk for people to walk so you can see from this image to synthesise supervision the generator is the South Action Alert or this Kaido on completion of concepts to you can't go back into the late and reprint patient identify them then we can control the different companies and all the Visual Concepts.

[00:14:43]
Another virgin point the importance of innovation in the people generating model I was seized late last year why such has so you know he forgot that we have this at the if he can and the p.g. get him to the Nathan bacteria He's actually at the big issues so the Such as he is as a beginning and then we go to sleep and back turn it into a generator that we can I think decide the image that there is a recent small those style down on the scale can be to the you to live the studio Why such as I think those who want the because you know vision for generating more buildings that's pretty city we have this treatment for what's at the beginning of the 4th downtown mostly have the skunk stands of the picking they just ask new anything factors through you to the years so each of the year they will add some even bacteria Sisyphus down through this type of things and they are wise to test the actually encouraged this the scene Handelman across the layers to right now we try to have her on the stand.

[00:15:48]
Today years captured on the visual information than we apply our from work to the later wife's Nathan doctors is actually we we discovered that there is a very clear semantic Iraq he's emerges from this thing to the side so after we apply our messenger to each may turn on back to recite the 2 years we actually observe on the 20 years correspond to different novel of semantics to the prologue on the last series that's the importance of the chillier the horizontal axis.

[00:16:26]
Mirrors that's the purpose the closer to the huge the various classes sure you're going portends to some level of semantics the right all we have this is a outs on that go object label attribute a level color so you can see in this. The generative group an Asian to Arlington a year actually come trolling the scene a outs to the layout of the things then the meat on your country road the object and actually suddenly through your control of the Congress so he's a roughly corresponds to how acute men came to the unit you know when when q. men we decide to like choice some image we will 1st I determine the layout of the things that we will ask our outlines all joking aside now we also talk on a rice determined different texture salt to the simple rip current patient is also doing this in an assisted 1st besides the be all done determine objects than none that had those texture and the colors to here I show you some dominance not to reapply all are nicer to different layers on the silk here after we identify the the textures that come troll with leaders.

[00:17:44]
Early on New Years now we can control their layouts to here I show you this and they all it's control results. This is a result of the interesting because this these generate through a mo those who were just trying this more go on those on Piers they thought they were treating the house just as those individual by the rules that don't know the from the rap so we don't have the multiple views for the same environment who is just on to care data to this generally the more moderate just chitin are those to quantify reprint patients structure information your own or to synthesize a new t.v. can go back to 25 and he report the missions none can control this layout.

[00:18:28]
The result he's knowledge is perfect you can see there are some obvious facts like some down on the right some missed out the gist of that so we don't have Anding it's pretty safe supervision on this but this is the result already impressive enough enough because we don't have any external supervision on this a month we will be using formation now we can control this structure are they out so here's another one on the meter only years we control it they're all jacked so you know people are like this the living room and the bathroom so you yearly all taking sides differentiates the 2 halves agree like we can get totally move out of the furniture just like the past and we can change just a rule into a tiny room right we can just move out to the past I'm moving the table on the chairs done this environment's well become more like a tiny room to the center the preprint patient also doing the same interesting intermediately year on the road determining which objects to put the site down we can prove them across the hasn't had agreed boundaries on the objects inside to be also changed.

[00:19:43]
So here I would like to. Tear them down take a look at them now that are. Modal So here we have the. This Mosul forcing synthesised full face synthesised so we Ching on those faces now we have this stuff 3 synthesise more though so we also have this recent war that's what we want to control and then the pie a dusty Rick with the missions in the face can't so we have this war cost in the press guns out separate the nation space between the generators and the actual space so the front works the stuff is also bursting or we analyze the different semantics the relations between different official action fields we actually identified there are some attributes of the mole in Pan goats with each other as in the mole on for some attributes most challenging to manipulate so often we have been to 5 different subspace that control the face then we can doing all kinds top issue as you can so for example we can make the sky cool nurse by Pakistan High School Aussies and or to make the latest in a the younger set and make this a person like from things the result was so interesting because we're trying the space gun we don't have this a paralyzed data to this motor just stuff done to fight it from the views of the persons that we can just push on the slating code to them change the abuse to kill him or to make.

[00:21:20]
More men who had showed just too quick time on all this. Video to hear is this time a mediocre Negra we manipulate the faith here we are in we just laid him codes and we can change the views of the the face the very least think. It's your own ending Cressy so far for this car.

[00:21:50]
So while I've taken a stand alone so good any questions I come to to respond. Yeah there's a handful of questions. That when you're introducing the topic. Someone asked how it's related to disentanglement Yeah it's a very good question and so does the same tango months you know it's just more like the emergent behavior you saw it's late in space right now we just to take this appreciated although we don't introduce an interesting aren't the same time going the mechanism so we just discover that the late in space we already seen tango itself was some of the attributes to some degree.

[00:22:33]
And so right now we we are working on some new piece entanglement mechanism through the carriage this process that we can a further improve this to manipulation a result of. There's another question about most collapse and. And maybe what you think about art I guess Jim Yeah the most most classic is a very common thing on me an awful can so you know the gallery is this diverse or trainings to sometimes do all of those some modes of the distributions that that's the this whole also happening in the be a you know variation alter in code or so so in our case we just take the pre-training that can to the preaching count him a job some mode but the skillful majority of the image even can synthesize high quality image that we can interpret their allegiance bait but in the United States we found out there are some on some peace force inside so they can space out from the wall every sample from the stuff they stand out prudes image will be quite bad so there will be some future worked out for you for the Star long come back.

[00:23:54]
Congress properties of the later space. Just we can there's 2 more questions about that part and then maybe we can move on Ok One is how can you describe how changes to the code are done during inference. So. During the different pollutant code is random goes hand is hammering changes added to bring factor to consider the object level to yes so so here we have this to mean there are some nations after we identify the boundaries we just change the meaning ask the mc just another of actors to we just add groups of actor on top of that there is a very simple process that you can check that our source code will see this process very simple then one last question it seems the foreground not change when you teams according to the boundaries.

[00:24:51]
How is the switch. In the foreground so that that's relevant so as we should novel we changed for those attributes out to describe some local regions so we can change the date and codes of the late 30 years the Army Ernie later well control more like global You fact the new image of the leader they are the country more firing that you do then we can come to add to the base for leaders.

[00:25:23]
Ok. So let's continue then we move into the inverse again so you know previous work they just work on the same size but in practice we want to work on the real face right so we want to work on my one face so here is my profile photo many years ago you know I have more hairs and I'll still gave them my previous profile photo so right now we have this gang version published so you tell universe for general control so the key thing this nation coats we won of giving this image we want to invert facets of the Nathan spent right previous generation it's something from the late encode to down time to restore a super sized image from the we have is the really mature we want to go back into the nation space identifies amazing codes that we have this debris inversion problem for them to stick this ally you believe down we can simply scold this minimization problem you know they think that we can simply all can my display they can go to them we can reduce that these terms between the giving image x. and this synthesizer the image that seems quite a unity right we just open my this now and I 1st tried this down in version and the so here is my real face I still use ice and so right now I get the Sultan my code after many to reshow all the documentation.

[00:26:55]
So let's take a look at how this I gulped meditation work so he's actually seems aside from the image that this through that he's reconstructed face so let's put a face on side by side with my own face so you can kind of see this minimization problem or these to really try.

[00:27:17]
Hard to try to things the size of my face but clearly it's not to me it's more like go was cumbersome so you can see the color had sure seems the same but that doesn't if you have the person just and when I go west I'm probably all me so you know when something doesn't work in machine learning all the people running we just we should try to get to hear I try different rendering initialisation you know this we actually have b.p. stock on this opposition problem so here I tried different runs in unison of the issues in just a synthesizer old different version of me.

[00:27:53]
Also you've clearly still done some file was. Organizing the more they will not generate your model because it's generating although they just train on this sort of bridge executing the stats listed us out on income town the Hollywood actors trees so you knowing this to Hollywood I'm not so many Asian actors actually so they cannot understand the sad truth why Asian people.

[00:28:20]
So so we want such **** if you shoot series discounting version is actually quite challenging proper So we want to fix this. So you know previous faults line years ago there is this mess are called Image to style gun you know I was a previous test there is a style can do or use monkey pointed him coat to reconstruct the improve the image done this master the action all can mice then they can vote at the different layers and then that the mission id calculation space is great to be increased so this kind of to my vision.

[00:28:58]
The the synthesised and the reconstruction quoted he's crazy improved so you can see here this he means to suck I'm kind of work so you will use a lot surely some space like a floating pie phyto on dimensions so my status Conry construct each. However there are some overfishing issues with a report that you know here is this general returned it's purely.

[00:29:24]
Human face to here I passed on some long space the image here I still found on tap the images site or some bathroom it incites to do this new version you just heard out seems to style gotten can you reconstruct the units this is seems quite amazing that too good to be true right so you can synthesize all kinds of long 15 minutes when this generated doesn't even seizing the training stage.

[00:29:53]
But this new thing that this process seems just older this talking the image of here we have some of them defies overfeeding issue we found out that the birds at the latest poll from the stimulus to Stoke on the cannot possibly all support the latest of many predation we just introduced in the 1st session after remarks and they think coats that we add some boundaries on public debate and codes and we can all.

[00:30:24]
Achieve this a realistic manner predations So here is some examples so we have this image reconstructed face and we add some a few mouse that's Tirzah all plotted factors though the quality of the read is that which means the new verb to coach doesn't very well supports this to manipulation.

[00:30:45]
Felician has this open my vision. Can generate some Alltop though may listen close to hear this old meditation itself we can talk on construction optimization problem so we optimize this so we may I'm not some on Nathan coach indicated that the Red Point is I'm a outside of the all regional.

[00:31:08]
Space you know is outside of the the domain the new the outside go in a time space then we cannot apply the previous some information Andris to this ultimate own domain they can go to right now we want to fix this. For this can you version there is the fundamental issue is actually the can use self laxness to infringe the ability to know.

[00:31:36]
They have a certain code or forget their you know in code or to over the years people try to add in code or so they can generate or you know take this approach change generally turns it just train some coders who got all the previous work that's doing the similar scenes like the pike in the Add some in code or through the.

[00:31:59]
Town models them to come to this future earnings also this year last year start the 1st 3 days now to include the results of the consumer saying this was a reason they wrote We have this one true coming quarter. So you can see this reconstruct the results from the encoder is seems an awesome boots on the quantity team soon proved over the years but the resulting in August still not good enough you can clearly see the identity of the request rocket face is not the same person for the inputs.

[00:32:31]
So here we have just a recent works and you can see we worked last year that's we try to come by the scene Khoder Nasser ways this scope my dish or masters to it proposes the code or constraints of transition to previously we have this to unconstrained versions on right now we have actual terrorists inside but here we simply and treat the 2nd quarter as the bigger riser for our under construction staff do you think if tied to 1st the reconstruction we also have the 2nd terrorists that's who we want to construct the leasing codes back to through the latest But you know for this on things the size images that we want to put into that include there's then code there will recreate a system they can code packed into and they think it's 6 to 8 after the 2nd person as a recliner for this these 130 the nation code tween force to make them close to accidents or regional nation states to this kind of replies or work quite well with Who here is a comparison with Siemens who stuck on model so you can see the reconstruction results improved also more importantly the manipulation results is also great city improved you see here we can and manipulate different attributes for the same person and so here is the video for the many pronation results so as we hear it we can change different expressions and those to change the post real persons run all the person a real person compared to the previous 2 and it's just the season size we thought face.

[00:34:16]
Unsecure we also have some. Interpretation results so giving 2 faces we can do this very realistic interpretation after reading perched on the poster late in coats of downtown just the way to some for the latent codes to come by the 2 faces. In small interesting that we can also do in this thing that operation for here we compare our results with the Nathan codes from our universe in the coats of trying to stop time so we count much better results and can generate some intermediary my computing style between the 2 also the result is much better than the I'm just stuck on.

[00:35:09]
Also we can include some new or pick a shame based on our mass so it's called semantic diffusion to more like right now we want to take the fall poke around face and in the background things that we just copy and the past and the stuck on that's even Xerox equal can pass the results read it that the what about doing that we can include system migration you know we can put some mosque to hear this an indicator not to right now we want to.

[00:35:43]
Keep this mosque the focus on the mosque. As the same as possible then just tell you not to pass along adaptive to the folk around than you can consider this as a diffusion process that we want to preserve the program and lots of folk around the information as if you still cross the background to after this title so maybe optimization and then we can make this pack ground out not to be adapted to the forefront Duncan call creates to serve very Clarence emerged.

[00:36:18]
So here is some time in the past wrongs the right to copy in the past and doing this opposition then we can have a very coherent program the backgrounds the size and shape a we're also adopted to the foreground so you can see here we have this young economy and Lisa in the my own face.

[00:36:40]
So I really like my this long hair style you can use this to change your own style that's why I. Just grow my hair salon to have a long hair. So here we propose some salmon. Because due to the time limits I can not to wrong this time and so we have this quality so you can go to the collapse and after loads of your own image now you can try different kinds of hairstyle different pants on styles.

[00:37:12]
Based on this semantical diffusion we also can use this to do this thing when I was Asian you know we can we want to like change a new bats in this things that we can just copy and past some padding inside and doing this in Moscow you don't need optimizations on the scope around well or the stock wrong well be adapted to the programs and here we want to change copy on the past new window and then to go directly.

[00:37:43]
Adapt to tag along to the foreground and form this very question here and each. Ok feeling always getting to come too close to the underdog talk so here I would like to promote another recent the library release 2 to fetch any page to research for general morning so here I call it Gen Frost So he said generally more than you for everyone so the so much fear them have implements for man high torture libraries for this generation modern history that provides many many different some of those who have them or those who are more than 60 preacher more those with squalid lives Thomas also provides the training codes for the Pi scorched for force on stuck on stuck on the 2.

[00:38:34]
So you can see here I generate some very crude visualization for that in compilation so I really like to skydive incorporations to hear small windows from a different model so I can watch this kind of interpretation for how they deal with this tell us that's different modal they have a different semantics so our previous or Ok You know we just use this off the shelf predictors but then people like the listen space but for this difference like any major animal it's we don't have this.

[00:39:08]
Training it's a class a farce Now how can we heard identify the report you mentioned seeing the sun long face non-value images polkas all those so here we wanted $75.00 to report the missions for different and more those who have we have this very recent work that. It's a Masters out doing this song surprised discovery of steerable dimensions so we utilize that we're looking to reprint patients to of the generators and try to factorize the meaningful and dimensions from from this a reprint patients to based on this work we've used some interactive.

[00:39:49]
Interface that's a user. Hang tracks resist the charity mojo to do this the collaboration that interactive content creation so I'm going after these is the difference I think the price they were going to mention the sound we can change the pose of this and he made characters' send them the size and shape of the staff and the nation so he's very cool read us that we also released the code so you can check it out on bail self want to play with the stem.

[00:40:18]
So here is some of it on each month so we're working with my very amazing student teaching so ramshackle minding how we also released this hour of our coats and the papers and Jim Fotis page you know he's a research initiative on general commodities because that really could be freeing up our all generally morning so here I have a few locals here who may generate a force be with you because I believe this to generate a more than can really improve this feel with all the competition Ok so maybe generated both the review through that and all my.

[00:40:57]
Talk so I'm happy to answer any questions you may have on those group. There was a handful of questions for this. Part when you're introducing the topic someone asked why can't you use your friend in quarter to find the factor of your face in the little space which is similar to what you did.

[00:41:25]
Using the quarter as a regular riser but you just framed the encoder itself without the other term in the law. Yeah you read just to use a in code or to reconstruct it that's according. To the quality that actually. The encounter itself cannot reconstruct results very well so here I show some in code or are installed on the reconstruction from the Iraq outputs of what you call their.

[00:41:58]
8 so this is a reconstruction also Gary $1.20 of the encoder although we have some improvements over in code or does this resulting in all I still not that do this also make a lot of scholars you know because even this in code or accounting reconstruct the Umich perfectly reasonable like we have a perfect image compression master to give me an image we can compress into a very short today some factors that we we can just repeat this jetpack o.p.m. to come crisis.

[00:42:29]
When you are showing some of the data sub bias results someone asked for the reconstruction here in inversion Kospi taught a visualization of different priors and vices. Yeah it's all very interesting questions so it seems easy there is this to talk to the boss so still all and even though many inversions he can call home overcome this pile see what I could become Bono's some odds are creatures to fear in the gap to reconstruct the content.

[00:43:07]
So that is my take away on this public deal with this stuff I always see issue in the challenging Moto you so worse exploring you know we have some recent work that to analyze this a fiar miss of the deep generated reprint Asian to find out for you fans are a lot of this algorithm a compiler seeing the generators so there should be further work around to call this eruption.

[00:43:33]
Here's a question do you find that the leading codes of cat lasted through her significantly outside of the priors for a network trained on faces. Yeah you know we analyze this so we it was like to test me for sure they should i go for this a face generator we sold all hope that the religious sites are doing this reconstruction and then we come by this week or regional Nathan coats then do this hits me because radiation so we have some knowledge of the Clearly form like to Klosters So we show it out so that you vertical Nathan coats using the outside of the region are neat and don't need to that's why off the rings just include there so we can sort of like push them all more closer to chuggers you know.

[00:44:20]
Make something so the next question is is there a class of priors for example counts and being imposed on the encoder wouldn't one want to ensure nothing covered South covers most of the truly distribution and doesn't suffer from some kind of clubs or. Yeah that's so that's a great question so for the call some prior season he's a very important priority so actually and then you get stuck on the you this culture prior to get they have the initialisation to see Michelle is the coach they did just have this a go shimming.

[00:44:56]
For the show your system made some offer we have the familiar white reprint patients you know the stuff they have the you should do some stock on the have the introduced some 70 layer is all for the Connecticut they are to further have generates intermediately some space so after we stumble from the couch in the abuse they will throw the latent codes into next 304 they are so funny coming clearly than they have been for media only 10 space they call the topic for this opportunity time space you should be ocean is on the No one thing in the case we can not to use and they have a need to call all forms to form the surprise distribution then we have here we just use a closer to approximate the Intermedia dating coach so that's can help out.

[00:45:46]
The output of the chinning coder Well we all screwed some nation codes that are closer to the same time medio days in space. Next month so the next question is what is what's your outlook for these methods to be applied to other kinds of art form and emission music.

[00:46:09]
Yes it's a great person so the 1000 for all my reason the work to do this so you know we have so many different. Images to different our arts and the open case and also the general theme of have been adopted seen many different made and I think it was a generation I'm talking generation 30 even can make a fatter cozier and that 25 year the same tango it's all variations actors don't care that's or Schiro control this process so that well that's people who work with this am although to do this a collaboration I go to makea I collaboration this is actually one of the future direction I'm pretty excited about like how can you not come work along with that I am also to create some new content not on limited to Umich also limit to the og or this acrostic of the dies or the logo the guy recent It is just the opening I dissipates to the medium although they provide.

[00:47:09]
Some of their options you know they have this taxing to remit they have through the generations but still with this kind of extreme it is still gnaws to like a steerable enough so we can recount Patterson to preach the lessons space and we can pass our customized the output we want now we can realise this to my credit ration.

[00:47:32]
Would be very nice. Where do you think he's fallen short compared to Cannes. Yeah that's also a very great question so the v.a. used going Turmel's in term of the McAuliffe He's a big auto school with the as Dan But the reason that a p 4 star like used some tricks to halt some some mass or to improve the v.a. like they have this movie to be a year that's the day you come tiles and they can space this action a very interesting Derakhshan they use the quantized ones that late in space previous you know the nation's space is just a continuous space they force some quantisation you cite this latest phase you existed Dr quantization So he has some idea from long time ago ng image processing they want to pass or generates a coding score they need some space because then they can use a quotation for formulas that.

[00:48:31]
Put some transformer size if they can space to pass on a structure like this latency than they can improve the. Image quality so that it actually gave us something that outfit we want during improved the demo those recounts a formulated some some structure so they can space then we can and do have better image quality interesting to see the future come versions of both to be and again more like a unified model.

[00:49:04]
With respect to the 1st half of the talk I'm any attribute class to train are the many wondrous rest class hires or is it 11 big multiclass classifier I'd imagine it's the former Thanks yes great so that's an that's a good question and so for our things it's a scythe.

[00:49:24]
Applied thing classifiers on the thing actually for the cost of power so the same process I classify the Occupy the cost of our actually derives from Professor James he has sung the sing attributes so there are there are $102.00 attributes that we're just trying to build cars to parse for each one of the tributes that we have $102.00 attributes plus a farce and then we apply all of them to some models but the we have this verification step so we found out about 10 attributes will ally with some of those because you know it sounds like in those things they don't have this open the.

[00:50:06]
Author is an exciting science attributes to the ending and there are boats like top 10 now tributes will have very well control some more so for different models the politics of that your boots will be different I put it all those things also attributes Well the minute I'm all for that.

[00:50:28]
But you will be well cared for about the time that you've got to have steerable in the control. Of the face and although we have about I think in our paper we task on going only about file effects official attributes but we can continue to add more but here we just treat each attributes as a binary attributes and there in the sense are there also some attributes that are not durable.

[00:50:56]
Yeah we found out some such a beauty Nah that's the ripple or so the nun knows they're going to turn. This into tango with all the action beat and we found out like h.. H. and gender are complex and h. and the Colossus is also a couple like increased the age of the person that is there will be a colossus that show up with also makes sense because you know for the other people in the world we are the glass so they.

[00:51:26]
Have this couple the attributes just as we did more disentangled more hope we'll be able to do with them even better interest Yeah exactly Ok so that's all the questions imo just missed one or 2 but that's all the questions I see right now. Ok. Thank you very much and I think we can and the talk that great thank you everyone for.

[00:51:59]
Advice.