Today it’s about Hitchhiker’s Guide to generative adversarial networks, very fancy. Name again, they seem to be very popular this year. Also, but let’s get started, let’s see how far we can understand this basis. So most of you might have heard about this. When I was for some semantics pre, we work with e-commerce, focused companies.

We do a lot of machine learning, algorithms data primarily and also intelligence layers. On top of them, we help them cover a lot of automated tasks like categorization brought us matching automated entity, recognition, proper products and, of course, the whole slew of offering this product matching distributed. Calling is one of our strong points, but I guess this is something you might be interested in.

The rest of the talks are committed to it. Don’t worry so, okay, let me give you an overview of how today’s going to be structured, we’re going to start with a generic overview of. Why do we study generative networks? What do I mean by generative networks and what are the things that involve, and then there are slightly theoretical part we don’t leave. I think it will be little bit of math, but should be interested for the fundamentals of how adversarial networks are set up for how brands themselves are structured and introduction to the system.

And then, this is why I didn’t want you to leave, because there are huge developments in the last year and then I’d like to motivate them through applications of how they can. We really visualize, and this should be a whole bunch of demos and seeing that it is a guide. This is like across the whole debate. I have links everywhere in the slice and then hopefully it points you in the right direction and then finally ganzar not going to be the most promising link queen Ganz are very promising, but they are going to be the solution to everything that you see.

So, let’s look at the issues which are commonly faced and, of course, again, improvements and the path forward. Moving up to the research is being conducted, the mic still works, so the slides alone. Okay, let’s look at generative networks right, so what do we mean by generated networks? So let’s get everyone active here and you thought the left. We have a think is a volcano. Everything on the right side.

It looks like hills. Can you get Mississippi is like easy. I think can you guess which one was drawn by a computer and what was done by an artist like a real human left right, anyone artist on the right? Yes, that’s Marvin from Douglas Adams. The Marvin also points out the artist on the right. I think sort of you suggesting that it’s too abstract, I don’t get it not very artistically inclined, but he has that the teapot on the right side yeah.

So that’s the boil something harder. Maybe so again, there is one of these was drawn by the bot and then here I thought of new Strax. I don’t remember which one was which so can anyone say left right, left right, computer on the left very good. Well done again! So that’s again, some study, which was conducted and it turns out people are quite difficult at making a sound. So let’s go to insane mode, and at this point I don’t know so both of these look very similar fancy.

One is like a very expensive painting and the other is 30 seconds on an Indian processor. So let’s look at this and then turns out it’s the one on the right and then so that seems to be so I here. If you look at most of my slides later, you have a description to whether sites are hosted it’s on an website. So it’s just click, so you’ll be able to see the papers which are referencing this and then this very recently came out in May this year.

I think El Gamal and it’s group they together MIT, is study and then a lot of people voted on the wrong paintings. So again these are house generated networks are structure. You essentially want to be able to do forgeries of high-end artwork make money that way. So why do we want to study generator networks we want to understand and that those complex information right? We want to be able to study, not just classifiers, not just discriminators, not just acting performance of numbers.

We want to be able to understand if there’s any other language behind them and then most of the time there is like high dimensionality in the objects being studied. So the probability distributions are no longer simple of one or two variable whatever we wanted. I mention it. What if the dimensions are in the hundreds or thousands you want to be able to positive correctly, and maybe that helps us finally, like in the previous example, we don’t want to just model them.

We want to maybe generate additional samples based on some criteria, and I think this title to the previous of themselves on reinforcement learning, because one of the proposed uses of Ganz is to augment the world environment. So say you have a reinforcement, agent and the agent starts learning. Maybe Ganz can help fit in and simulate the environment in which the agent learns. So they pretty much tie in together quite nicely at that point, so generative networks so good to study them, and then let me just motivate how they are going to be structured right.

I’ll take a very simple example: there is me xx very messy very soon by then. Let’s take these points, that’s a whole bunch of points on this distribution, and then these are the points which we want. The model think of these as data points to our structure. Now one way to do it is something called very popular. It’s like maximum likelihood. So wherever we see the points being recorded, we adjust our probability.

Distribution function, sort of upwards. So later, if you look at the blue line, it sort of models how the points are distributed in that particular dimension, and this is very common – this a subset may be the most popular one of how generator functions are formulated mathematically. You can define a probability. Distribution function, P and then the P is based on parameters theta and then you start describing the speed of model which gives a sample X conditioned on or is greatly dependent on these other parameters.

We want to optimize these parameters, do something like the actual operation, where we do gradient descent back propagation of those things, but basically operations where you want to optimize the parameters theta, and we want to get to be right and from P, we get this distribution and From there you get the point, so there are two approaches: P can be either explicitly defined or we ain’t be carefully.

That’s the one on the left hand, side explicit density functions, though, is that they’re like a whole way of approaches and then over here as a root nodes, you see all the popular approaches I just focused on a few for now, so that you get an understanding Of how everything is set up, so when you look at a operation which, like eggs, to get the explicit probability function, you can either do it as a retractable method.

So if you look at deep mind in the last one or two years, I think they came out with wave depth wave. That was where you enter text and then it automatically speaks in very realistic audio. The person speaking sounds like a serial person speaking and not like your old Microsoft thoughts they could now but supports which that useless, literally promise word by word from recording. So we’ve met one based on this sort of approach where they had fully visible belief, net, which are very popular wavelength.

But the problem was that it’s based on, like the probability of conditioning on every single input, because, like the sequential learners which was expecting to model the explicit density function, another very popular approach is the variation auto encoder, where, instead of aiming so, we don’t want to Get the perfect p, we set another function which is a lower bound on the. We call that maybe L and then that variational bound is sort of proven to be optimal in some situations, not at all.

But the idea is that you sort of optimize your function and then variational auto-encoders are at that stage where they tackle the lower bound would be existed. Entity function, that’s the left, 3/4. Now, on the right hand, side again, you have also changes which are very similar. They do repeatedly sampling and they’re two approaches to it, but for now F be the end vation auto-encoders. They are very popular because they’ve been giving very good results as, for instance, they tend to be functions.

What do I mean so? Instead of assuming as a P exists or itself aiming to get the P, the probability function, we just assume that it exists, so we say that okay, there might be a fee. I don’t care about the probability distribution, probability density function. I instead just try to get the samples out of it. So again: yes, the gastric networks, markov chains and then it’s that aspect. So in that sub-segment of the tree, where we assume that all the identity function and then we start generating the outputs of the model that sort of square or both south is going to be centered on.

So we’re going to look at ants in this context. So if you start from generator network in general, Gann sort of written displaced relative, one of the probabilities in the probability density function so many times things just keep coming and to say that, so that’s how the ants are supposed to work again. He and Goodfellows very popular with the tutorial is there on the link. You should check it out it’s much more than I can fit in forty minutes.

So how do we do this comparison right? So when I reported wavelet, which is like a very popular model, unfortunately, it takes two minutes to generate one second of audio, so it’s like very difficult to you can realize it’s very popular gives very good results, but the computational complexity is sort of like a problem Guaranteed asymptotic consistency, although our main methods there, of course lower bounds and you expect to optimize it in terms of auto-encoders and bae, they sort of don’t give you the full result.

Eventually, you hopefully get there, but the guarantees are not a hundred percent and Markov chains because of their consistent, resampling and continuous requirements. They sort of take a long time to converge. If you’re play with Markov chains, you realize you don’t know how many games to play of the multiple iterations, again rate description, links below the slides are not exhaustive. Also, these are also based on perfect sort of existing results and dance again are not going to be perfect.

This is just an example of why can you not suffer from these problems but later on, towards the end, we’ll see the problem that they have of their own? So now we can jump into it, generative adversarial networks right, so just to give you an idea. This is like the statistics. Over the past year, it literally said cumulative gain names against people, sites calling names, sim game discs, organic slice against PC, get so many games right, and then they decided, let’s plot a chart that even after 2017, you see this right here.

It’s literally like a zoo like so many abbreviations that people have no idea what to name the against. We are running out of abbreviations for again, so if you can think of one make sure to put a preprint and Rainier span. So let’s look at the gandu. Now, of course, was very interesting that last week was the computer vision and pattern recognition conference and turns out Apple started, publishing their own research papers, their first one, which came out, which also won the best day for our last week, was on gents figures cents.

So they started generating eyes for others, relations and it turns out the leaves are quite effective in training their own models. So things moving very fast, and I guess before I finish my thoughts, I might be another three or four preprint. So let’s look at the answer. More detail, I like sex, like just talk about the structure, how they fit in a sort of training, is involved. The easiest way, like I said, is adversarial networks.

So there were, we do players over here by players. I mean eventually two models: dude, your networks, two systems. Currently we just give them as column players right. So that’s a tea energy. The names has become apparent very quickly. Why sorry just check? Okay? Okay, now it was maybe it was not selected, so the adversarial gain can be formulated with like a game theory formation. For those of you have an idea.

It’s like the national system between two different agents that the local equilibrium, hopefully between the two, which sort of arrives at a stable solution and then the two names are for discriminator and generator. So one is the forgers. The other is the expert, and then they generally are playing together where they are trying to compete against each other. There are also formulations where you can think of it as a cooperative game, but so far the way of think of it as like us, agent, white tooth peoples, might against each other sort of skills are easier understanding and it me James to defeat the other.

So if you think of this framework, let’s look at this picture again, so this is pretty much how you can explain how a can is structured together. So let’s look at the right side. First, where we have this sort of noise, so you can think of this noise as the vector space. Let’s just call it a latent vector space and I’ll denote it by a vegetable V and then this V sort of fed into a neural network, the generator and then the generator has a function.

Let’s add a G of G and then it starts generating samples. So it’s just start generating one example of another based on some random price, at least initially, so it starts generating output on the other side over here you have a sample data set, and all of these are real light. Images are real life examples, so these are other fake ones, but the real one, and then these real ones are going to be where we sample our true values.

So what happens is that you have this network at the top? That’s our second network that will be called our discriminator, then very simply the discriminator is either fed real values. G of X is the function then or it’s pet output from the Jen that becomes D of G of Z. So you have the effect. So G of G of Z, one of the view feet one or the other into the discriminator. And then it’s like a binary classification problem and I just raised it the real image or it’s a fake image.

It keeps getting again again and then these two systems are what we are going to frame now. One of them starts generating the other one starts discriminating. So let’s look at the really easy. I want to think office or the discriminator, which is one at the top. It thinks of whether quanta to determine per sample is real or fake. So when the sample is from the real set, it wants to say it’s one where the sample is from the fake set, it sort of starts.

This function enter zero. So that’s like the complication of how it goes about it. This is how we want the discriminator to behave, and then this is a very fancy way of saying the two lines here. So we define a loss, function, the J for the discriminator as a function of its parameters, theta and then, if you just close your eyes or squint real hard, you can just see here where T of X is the ones which need to become one and D Of GOC pretty much in a 1 minus, so the idea is that by combining these two expectation values, we now have a easier way of defining how the log function needs to be written for the discriminator.

So you can think of this as the final loss which is going to be calculated and back propagated through the network or the discriminator updating. So we have the discriminator, it’s quite straightforward: 1. 0, it once discriminated, then we have the generator for the generator it. The idea is to put the discriminator, so we can sort of use like the recursive thinking about it. What are the send winner? It needs to pull the discriminator and the simplest way is to just define it as i-.

So you take the discriminators function. You put a minus sign, you call it for the generator. This is sort of like a min/max game, where you make sure that one of them is able to defeat the other. There are a few problems with this, because the think of them as two separate networks, so, if that this later becomes optimal, then the generator stops learning because one of the lofts becomes zero. You put a negative sign: doesn’t matter zero, so you sort of fool.

The generator is thinking that it’s no longer learning. So there are some heuristic motivation to get another way of defining the last function for the generator, and one of them is this. So, G of Z, that’s the output function for the generator, and the of UFC needs to be one over where this is different, because the generator wants to fool the discriminator. So he wants to convince people that becomes one.

So you write like a non saturating. Heuristic function – and this is pretty much to say, is that the expectation needs to become one for the generator output. This is how the generator thinks about it. There are other ways like I said this is formulated in a maximum likelihood type of environment. So you can pretty much pass the equation where you sell here, like the logistic sigmoid function on top of the output.

This is just for defense, so that you understand that this is actually functionally equivalent to the maximum likelihood estimators I was talking about, so this is pretty much just a way of recasting it. So you can think of this as like the two popular ways of how to define a lot function for the generator. So now we have 2 log functions, one for the denominator, one for the generator. What do we do? Stochastic gradient descent, most common, so we take two mini batches.

We take one from the real samples X and then we run the vectors on the generator we get another batch of output, so we have too many Westerns one real one fake. We pass it through the discriminator it gives you answers, gives you ones and zeroes and then, based on that, you can simultaneously run back propagation on the two networks, so you calculate the loss for the generator you use it to update the weights for the generator you Calculate the loss for the discriminator you use it to update the weights for the discriminator the setup with each one, whatever loss function to calculate can be used to update both networks, and this sort of joint training is why it sort of pillars together.

So, for a long time, so this was back in yeah. I don’t have a theory, but this was back year 2014. I think when he had good fit on the other sort of formulated material system and then for a long time, there was not enough for successful computational EDC implementation forever and something is called lab. Ganzalo prescient betterment operators and using that was slightly difficult and the whole thing didn’t take off until much later when we had our familiar congressional networks and deep conversion networks being applied to them.

So this was sometime in 2015, when a group proposed that you take your layers, you make them all convolutional, remove any pooling or run pulling that you have and then eventually you start to expand, generalization and turns out. The output of these sort of networks was much more clearer, higher resolution and also easier to Train. So just a few pointers here. So what happens here is this is the network of the generator.

So we have two networks. This is probably the generator. So you look here: it takes like a vector set and then it has a whole voice. Job for convolution excited convolutions over here it makes it bigger and bigger and bigger. So you can think of the final output as like an image which has been generated. So this image is then going to go into more traditional classifiers and then that’s the one with such distinguishing, whether it’s a real image or a fake image.

So this like the deep convolutional again which were set up in order to successfully make it work together. So that’s about it hope that wasn’t too much of technical detail. But now again we look at very interesting applications. So with that setup of how generated disk invaders are now working with each other, you know an idea of the different components involved and, let’s just see how that translates to real-world applications and the first thing, which is also from the same paper, which I was referring To DC gas, and when I saw this, I was like blown away for a few minutes.

I had to sit down and think about it. So what happened was that if you remember what we went and all the complications that are all sorted, we look at the example and you see very strong similar. So when the generator starts generating output, people were able to model the ways in which the outputs are related to each other. Let me give an example here, so there were a whole bunch of pictures. All of these are generated pictures.

None of them are real pictures. So all of these pictures were generated by the generator or a specific position in the latent vector space. So you has a generator. It takes some noise type of input and then gives these photos of men with blocks. Then for some other point in the vector space. It’s just leaves photos of men. So what do you do you subtract it out, and then you can see where this is headed.

When you add a woman with glasses from the vector space in the noise in the latent vector space, the generator seems to be able to transform it like this. So if you remember the words awake approach where King Midas screen plus man, that was here, but it’s out of the way where you start seeing that the vector is able to identify the important components in each base – and this was something that’s quite interesting.

It’s not just segments of the image or it’s not just lines and edges which are being identified, but it’s really the concept of the image, and this seems to be quite interesting of how the generator networks are learning. That’s another research paper. Most of it is going to be this so just assume that the link will be there, so you take like a very strong image or very high-quality image, so this is like a very good painting.

I guess I think I think editing the painting and then you down some to the painting say it’s a thousand five thousand pixels you make it much smaller, wonderboy, Boris, exactly and then you put into Photoshop, and then you run the bicubic interpolation or some simple method. I hope it’s appear, but this is very blurry, so this photo turns out when you approximate each point by the neighbors, you get a blurry result.

However, it is possible to take this sort of lorry down. Sampled images run it through something which is called an Ed’s organic. So many name again, so you run through it again and then it’s not generating much more clear images. So it starts giving high-resolution images which sort of end up being typical improvement over what was previously proposal. The super resolution seems to be possible now interactive Gantt.

I guess all of us went to school at some point in it. Remember those is art, class and officials that you draw a painting and everyone goes. The scenery draw a little dot. Two mountains draws Sun, you’re, really fancy you draw birds and a house in front of you so again looks like the gantt are reading at meeting us at it. So over here, what you see is actually a person conveying his intent. A word like a vector space over the noise vector space again and by simply telling them what colors they want.

They’re able to do things like real-time editing of images, so by simply drawing a white line, he sort of adds no to the mountain by drawing like green lines over here it becomes like a like a field. The interesting part was that for each of these updates, it was of the order of seconds, so we are literally able to hear real-time edits of the images. So that’s what interactive again! I think there was another impressive article where they took a photo of a lady and they started adding black colors.

Then the hair change blonde to black again very impressive result. So these are like the rough images, but it actually starts generating high resolution images once related to settle down image to image translation again very popular. Most of you might have already seen. It also referred to as a fix to fix, and this is sort of what was being done. So, on the left hand, side we have input on the right hand, side, they are expected outputs from the generator again.

This is not a real picture of a street with cars on it, a person sort of describes how it will look through segments and then it starts modeling it and such if outputting images which are trying to fake people. Similarly, if you have spatial photographs there, you have a new satellite and it starts taking photographs of these earth, and then you start seeing growth and maps be auto-generate. It can go even further.

It can even start doing impressive results which are being done in Photoshop like if you want to convert from black and white to color. You want to go from bad dancing to a nice Street if you’re an architect and you design a house, and you want ads where the doors and windows are going. It may look impressive here, but if you zoom in there might be something very weird about it, but I guess it’s part of the process of how these registers learning.

Of course, you can do things like this, where you draw a rough sketch and then it starts giving realistic interpretation of how it might look in the given domain so create website over here you can go up, try it. They have very impressive other models, not just fiction books here, text to image synthesis. So a few years ago there was a very popular article where you give a photo, and then you start describing the photo automatically.

I think Facebook even now does it. If you have got a photo to Facebook, it starts telling you that scribing add keywords to the photo which sort of describe what’s happening there we’d like to do it in Reverse here. So what we do is that we start writing sentences and these vectors sort of condition our input. So when we feed to the generator we sort of add additional input into the vectors and then these sort of vectors, the text vectors, which are encoded as it’s not vector space, they start determining the output of how the image is going to look like and then Even for me, like I, couldn’t do this right, so you start with describing this and then it starts coming out with realistic photos of how a bird with a red V and with like a black feathers standing out.

There also possible very impressively for flowers and some other aspects, so this is literary translation in the context of going from text descriptions to being able to go to images again. These are again very terrific examples and all of that, but seems impressive, that it even works in the first equation: image completion – and what do I mean by this? It’s about corrupted images right, so there are whole bunch of celebrities here you might recognize a few.

I don’t recognize most of them, but you cut out their nose and then you train the network on how to do in paintings over these images, and then it starts to identify that adult is most oftenly the structure over there. It starts to match the color and the shape with the rest of their faces, and again all of these are ways in which the whole system becomes a part of it. So it starts to generate sort of realistic results and again generative network is the center of it.

You start noticing a trend. Most of the results are 2016 or 2017. So that’s like the last year. It’s when was the things that are happening and it keeps improving day to day, multiple Gant, which I think this 2017 now right now, you months ago, we’re back again and people are able to describe them as combining to gain networks, so, instead of having a generator Discriminator by itself, you take another pair of generator discriminators, and then you put them together and you start training them to identify relationships across domains.

So what happens here is that you go shopping and then you have a shoe which is in a particular style, and then you want to get a bag which matches that scarf. So they started being able to generate samples which are consistent across their whole catalogue. Very impressive results and then there’s a new one, so this was called disco pants and the other one was called psycho again. Don’t have to remember the names but yeah.

So there’s this also well here and they were able to like take a photo or a article of the horse trotting around and one network was able to do it in to convert it into a zebra. So when I looked at this, I thought. Okay, that looks like a zebra, but the trails pervious. Then such zebra trails are not striped, but I didn’t know that so yeah and then you could also do it the other way around where zebras could be converted into a horse.

So these sort of games were able to do it in either way. So you have tracked against all the cool games and then it’s just going to be all the way down. You keep going with gems, so very impressive results and, of course, a few more minutes and then we’ll talk about issues that were happening. What is the state of the art research and if you want to go to the nips this year, I think solidly closes the papers have been accepted.

But most always, I could end up talking about things like these, so stability and other problems that you just recorded, increase them. So finding equilibrium are hard because, like I said this is not a explicit last function, which is just being optimized, we need to be able to do sort of a joint system. We’re optimizing address one network, but together with another one, so the local equilibrium points are sometimes local minima, which you end up being stuck in non convergence, has always been a problem, especially especially something called both collapse.

So what do I mean by that so say? You have pictures of dogs and the idea is that you want to start generating fake pictures of dogs. You feel it like a Labrador of world or golden retrievers, and we start reaching with all these types of dogs. But what happens is that, as mentioned like the view we mean like the few talks earlier and some people think that the networks are of learns to cheat and it starts realizing that okay, I will just start generating bugs and those are like.

Well, it’s dogs go up, so it ends up that we sort of collapse into like a diversity problem. The network starts generating very, very similar-looking examples and, of course, there’s one thing which I sort of skipped because wasn’t very apparent over there, which is the differentiability requirement or the loss function and the whole network. So you morality of C R, T of Z – and you have read both of the either of the functions – need to be differentiable and go to much of the mathematics.

But the idea was that this constrains us sort of continuous functions, just being output, so working with text has been a problem in this area, so most of the functions are most of the outputs. When I notice a trend are all images, there are approaches which are looking at generating text with realistic results, but not into convincing so forth. There are both close, but let’s see, and how do we improve can’t so stability has been a problem for a long time.

I’ll just leave all of the tears okay, so the first view are on hacks of how to improve and the sewage container has like a very popular repository when he did this to it. You want the Train again, just follow these steps, and one of the idea is that, like the first vector space set in which you start generating the output, instead of having in a select an IV uniform distribution instead of having like a spherical space – and that seems To give by t better results, one sided label, smoothing which may be applicable to other domains as well, essentially means that, instead of going from 0 or 1, you sort of adjust your label so that your output is like a point 8 or 0.

7. This sort of still conveys the information that you are good, but sort of adds noise at the layer. Finally, you also have reference normalizations batch novelizations. These are something started by says: the losses are computed and the intermediate layers that organization has done slightly differently and finally, this is sort of like the big problem with Ganz which people are still working on.

It’s about stability in all of these systems. So when the two networks are learning against each other, they’re, not very sure, on how stable the outputs are going to perform, so people are looking at different ways: the expanding to load when almost all of them are sitting 17, but most of them are still trying On how to ensure that stability is a solved problem, one easy or one effective approach seems to be conditioning on how realistic the output looks so instance.

So how do you search the quality of a fake image right? Everyone has their own opinion. So when you start judging the quality of fixed images, people select defining it using other distance functions. I just messaged, I want one of them, so you start defining other loss functions. The other distances based on that, where you make sure that the realism of the image is sort of an important parameter and also Auto, is other approaches like the Egon and began.

It has actually work now so began and then these approaches sort of like help, define or help set the stage for stability problems to be involved quite sure, a lot more to be covered, but I leave all of these here, so that was the guide to. Can you can visit this link for the slide and but that and all the references, hopefully it’s useful? I just wanted to give a presentation on all the overview, maybe hopefully abuse when you start searching off the checklist guides again there was media Mona and thank you very much we’d love to offer you a few minutes for questions and answers.

If you have a question, please raise your hand we’ll get you a mic. Oh I’m going to write down here, I’ll, give you a mic um! Please stand when you give your question and I’d just like who you are. We are doing a live stream. Hi thanks for the interesting talk, I’m the food, so my question was around. You said that differentiability requirement is causing problems with discrete outputs, so that applies when I’m trying to go from say text to text.

So you can’t do my homework yet, but so what about other forms of continuous off fruits, for example? So images is one, but what about the graphs, or you know, time, series and things like that, so have they been areas like curve? I think that have been few approaches, especially like at the start where people are trying to generate audio based samples or even for images right. You can pick up it as so.

Essentially, the pixels are going to be numerical values which are being generated. So I would think that if you want to solve problems which are time, series or value based the same encoding system or again, as I don’t remember, of any specific approaches which I work. But if you look at an image when you actually type it out, it’s going to be a series of numbers, pixel values, so I think that might still work and those are not just discrete yet but quite sure it might work.

I don’t know. I can’t remember here of any specific examples. My next question is a winner hi. So when you showed the equation for this Renee, when you read it for this degenerative equation: okay, catalog, when you said it’s only finery, okay, what? If the problem is an equal person? It does it as a dragon, go match. Yes, so I didn’t talk about how the network terminates. So what happens is that when you start off initially, the discriminators has the option between zero and one, and the whole structure is based on whether the loss is like a finite value.

When the discriminators becomes confused between a half, so it says X is half that it has. That sort of the instability condition that ideally, when you start the training. So what happens? Is that at that point, you’re at the place where the generator is like completely as good as it can get, and then that’s essentially the stopping condition. So if you have the outer of the discriminator stuck at half or 0.

5, and then it cannot differentiate between the real and fake images, that’s at the point where you start seeing that you stop your condition and that sort of the end state of your training. Yes, yes, so the equilibrium condition is exactly that. The equilibrium position is that the disco mater fails to generate any difference between the between the two images, all right and ask questions right here in the middle hi yeah thanks Don.

So this is also strengthening stirring through the process the distributed effectively. Sorry, I couldn’t hear it quite clearly here: yeah, okay yeah, it was very interesting. Is it supposed to discriminate to generate your strengthening through the process discriminated the factory classifier? Okay, yes, for their learnings. For the effectively the architecture, the way the discriminator eventually strengthens from other classification problems.

Yes, so this was expect something that’s quite interesting or that you have two networks which are being that here and most of the results are focusing on the output of the generator. So one thing which might be possible is to take a very good discriminator, which has been trained using this process and literally start using it as a classifier or other types of problems. That’s definitely what motivated me also, because I’m more using the classification problem, so this discriminator by itself might have significant uses in solving problems.

It might not be trained explicitly on the classification problem that you are trying to solve, but again there’s definitely some amount of transfer, learning or now want of retrain weights, which might be useful, I’m not sure but hope, I’m quite sure there are. But I stuff my mind: I’m not sure yes, but definitely something we should look out for with question weigh in that most of the applications which we saw was from the generator part.

Is there any like? Can we use the discriminator part for exactly? I think. That’s exactly what the previous question was also about, so I think the utility of the generator is what has been discussed so far. I also strongly believe that the discriminator needs could also be used, and I think, most of the times when you want to use the discriminator it’s sort of not at the image context, but maybe in the classification context, and some of those are quite useful.

We are not seeing anyone tries to use the discriminator for likes, imagenet, glasses or anything like that. I think for simple binary operation of like deciding whether an images force of a particular class or not. I think this matters are still able to do quite significantly whether they have been used anywhere so far for any classification problems not aware of any okay. If sorry can’t hear you a really, maybe my people, oh okay, better later so this motion is no.

No, no, like you might have come across the different, fake news generation like Obama’s speech, has been moved. Okay, yes, what we can do so how ganzar like helping and tackling this fake news challenge, or are there any other any countermeasures to handle such big news? That are getting generated using ends like screaming, for speech based input for organic, because xbase filtering not aware of any good results because of the whole problem, with discrete inputs and discrete outputs effect, but for filtering of voices.

Other images like we can identify this is generated by some victor so that the fake news can be in the future. Possibly, hopefully you get a network which is talking. This involves using the discriminators component, I think yeah so in the decimator component, and if you have a generator which is aiming to models of the fake news. Part of it quite sure that, as long as you generate your samples consistently, it might be able to use the encoding into a representation of what a fake news represents.

Is the actual problem, in my opinion, so getting the presentation right might be an issue. We have sense of one more question and we have a question down here. Okay, one way you are talking your board or you know I explained about the thief and you can draw the is using jack, but at the same time, when I look at you yeah when I look at the disco gasps, I see that I, the heart, is: We change to G Brown.

Okay, I can actually put some Rafal some wrong person in the same article. You are usually in the judiciary system, be kind of relay the articles as well. It’s kind of regularity on this kind of mind. I research on the instrumentation and the research good question so far. There have been very convincing results. I think, a week ago someone made a made a picture of Obama’s speak very realistically things which he never said.

The legal issue substances a minefield. I have no idea and how people are going forward. There are few organizations which are looking at it. Efs the Electronic Frontier Foundation open a eyes to an extent. All of them are looking at the ramifications of how reinforcement learning systems Gantt systems generator networks. How all of these will fit into the broader context of how it might cost issues? We’re still not sure some people are convinced that it’s going to destroy the world.

Some people are convinced that it’s merely a pet project, but we still need to have reasoning with the entire society, no longer computer scientists who are going to determine these things, but yes, due to source of the skills for me to think about here. Yes, yes, this is such an interesting topic and I think we could probably have questions all day long Ramadan. Would you take questions offline during you, yeah sure I will be here the rest of the day I for a while and it’d be happy to meet you anywhere website, also fantastic.

Well, it’s time for a break before you guys go out to get your child off e or whatever caffeine. You need