What are GANs (Generative Adversarial Networks)?

Learn more about watsonx: ibm.biz/BdvxDJ
Generative Adversarial Networks (GANs) pit two different deep learning models against each other in a game. In this lightboard video, Martin Keen with IBM, explains how this competition between the generator and discriminator can be utilized to both create and detect how you can benefit from the competition.
#GAN #GenerativeAdversarialNetworks #AI #watsonX

[@baqirhussein1109]

I like the way he smiles and the calm talking

[@julesnzietchueng6671]

He clearly loves his job and its communicative ^^

[@ahmedaj2000]

loved it. simple enough to be understood yet complex enough to get the important details

[@KW-md1bq]

I don't think it's very nice to talk about someone else's amazing invention without mentioning their name. (Ian Goodfellow created GANs in 2014)

[@canaldot.5243]

Wow, this is the first time I really understand the concept of GAN. Well explained. Loved it

[@TheAkdzyn]

This was excellent. Came across gans a while back but some of the explanations i got were deeply technically complicated so I couldn't quite understand them properly but this was very precise yet relatively concise for the amount of information it conveyed. Well done. I'll look for more from you!

[@AishaKyes]

this was so easy to understand and interesting, thank you!

[@aryamahima3]

Just loved his attitude and way of explaining the concepts.. 😊😊😊

[@tanezcorvideos]

Really perfect explanation of GAN, well done!!

[@shubha07m]

Just one sentence: The easiest yet more powerful explanation of GAN!

[@deyon4521]

How is he writing with his left hand, from right to left and mirrored so that i can understand.🧐 Or is this just his secret talent.

[@IBMTechnology]

If you want to find out we shared some backstage "secrets" on our Community page, you can check it out here 👉 ibm.co/3pT41d5

[@toenytv7946]

Elementary my dear Deyon nice one.

[@sc1ss0r1ng]

He's writing it normally in front of himself and then they have mirrored the video, so we see what he actually saw when they made the video.

[@SheSweetLikSugarNSavage]

😆

[@recursosmusicales399]

Is a fake 😱🤣

[@suvidhibanthia212]

You made it so easy to understand. Thank you!

[@lethane11]

Superbly explained. Thank you

[@jayanthmankavil]

Thank you, IBM, for these videos!!

[@robertdTO]

Excellent, clear, to the point in introducing GAN.

[@nokostunes]

kudos for the clear explanation + writing all those diagrams backwards :]

[@vrundraval6878]

this is what you call a clear explanation, thanks

[@IBMTechnology]

Glad it helped!

[@usamazahid1]

elegant explanation .....great job

[@xmlviking]

I absolutely love this topic. The advances in human medicine could be incredible with this. A sample "input" from a bio organism...and then a model "of you're target cell types"...and then prediction on outcomes...and then further samples of "feedback agent" and then training you're human cell model. Then we introduce the GAN and think about our models accuracy. The future state possibilities of identifying interactions "trainings" with various drugs etc. This type of interaction could lead to identifying bio organisms not just humans and potential outcomes of interactions with them. Extrapolate that with humans and food allergies, diseases etc. It's mind boggling. When he is talking about CNN's and the use of alternate examples with Discriminators and Generators with Encryption my mind exploded. You could, hypothesize a Hedy Lamar like frequency agility but apply that to encryption and use an encryption agile chain. Good lord, super computationally expensive but man that would be nearly unusable from theft point of view. Would take you forever to crack that..as all the data could change from one form to another over time of transmission.

[@TWHICH]

damn

[@kitrt]

How far are we from networks that generate networks, I wonder. Like a network that tries to produce the most efficient neural network structure to achieve a good enough result in the shortest amount of time (or cloud resources) in a given use case. Or it's more efficient to just use genetic algorithms?

[@MasoodOfficial]

Excellent Explanation!

[@Surya25398]

It is really helpful, thanks for your video

[@yasithudawatte8924]

Very well explained😇, thank you.

[@somuchtech9864]

Very well explained. Thanks for sharing

[@mhmoudkhadija3839]

Very nice explanation! Thanks sir

[@syedmuhammadsameer8299]

For the image upscale problem, would we still feed the generator random noise or will we give it the lower res image?

[@sapnilpatel1645]

Very Informative video.Thanks for making it.

[@petchpaitoon]

Thank you, It is informative

[@Democracy_Manifest]

Great video, perfect presentation. Was this artificially generated?

[@sathirawijeratne7872]

Love this explanation!

[@taqiadenal-shameri3800]

Amazing explanation

[@saharghassabi]

Thank you very much... It was so intresting way of teaching this network

[@huynhphanngockhang5733]

oh i like his voice so much, he teach very very easy to aproach

[@gauravpoudel7288]

Appreciate the effort put into generating such great content. BTW I don't quite understand how generator and discriminator concept can be applied to : predicting the next video frame OR creating higher resolution image These were discussed in the video at 07:15

[@parteeks9012]

It can be used as a discriminator. As we can feed some part of the video and ask him what the person is going to do next? if the prediction is correct then feed more hard questions otherwise discriminator has to improve its weight.

[@user-xn8wg6yw7g]

Good explanations. Thanks.

[@engin-hearing5978]

Very nice video and super clear explanation. I would like to ask a question, staying on the architecture of GANs, one could believe that their results would periodically improve. If this is a possibility, are we measuring how much deep fakes improved from one year (for instance) to another? I think would be interesting to know it to understand if one day we will still be able to detect them through digital forensics algorithms.

[@Arne_Boeses]

With better and better Deepfakes generated, also the tech to detect deepfakes gets better and better.

[@reggaemarley4617]

@@Arne_Boeses But will detection technology ever be able to outpace generation technology? Based on this video is sounds like discriminator type systems are destined to lose.

[@GigaMarou]

well explained sir! but i don't get the application of GANs in the context of video.

[@andyjc9558]

Can I use GANs to generate a lot of Fake defects images of a product and use to train a 1st model?

[@animanaut]

what is the difference between a discriminator and a classifier? or are these synonyms. reason i am asking is: classifiers are sometimes mentioned when it comes to detection of generated content. but, if a discriminator in the endstages of many iterations is basically no better than guessing it does not seem a viable solution for this problem

[@usama57926]

good explanation

[@yuvrajanand1991]

Simply Loved it

[@user-bs4vu6mw7f]

I want to generate images through GAN from MIAS dataset. Which GAN architecture is most suitable?

[@aryanarya72]

I loved the way he said in the end - "turn a young, impressionable, and unchanged generator to a master of forgery".🦊🦊

[@betrunkenerbierkutscher]

Thank you very much for this video it was very helpful and comprehensive. ☺ I have two questions regarding the image generation. Maybe you can help me:) 1.Taking your example of generating a picture of a flower; does the generator have any kind of "knowledge" of how a flower roughly looks in the beginning? Or does it randomly give a pattern of pixels to the discriminator and learns by the rejection it gets? 2. How do GANs work in the text-to-image generators? For example, I wanted to have an image of a blue banana and my GAN gets this input as a text prompt, how would Discriminator and Generator tackle this? Would the input be relevant only to the discriminator? Thank you!

[@praneeththota5459]

I think I can answer to your questions 1. Yes generators learn to map random input vectors to fake flowers without any prior knowledge of how flowers generally look, however one can use a pretrained encoder from Image encoder and decoder neural network that has been trained to encode and decode flower images. This way the generator would have some prior knowledge on where to look in a given input of random vector to generate flowers thus making the convergence faster 2. In GANs just like how we pass on random input vector, while converting text to images, one can make use of an encoder network to map the input text into embeddings (something that's called word embeddings in the NLP domain). Now these embeddings can be passed to GANs inplace of the random input vector. But in this case the descriminator has to have knowledge to perform multi-class classification, as text-to-images might involve generating multiple objects/entities unlike in GANs alone where we try to generate only one particular entity like flowers, or faces or cats etc

[@sitrakaforler8696]

Dam.... thanks for sharing it so clearly !!!

[@Aimeecroft]

I dont know if your still responding to comments, but ill give it a try!. Im currently looking at deepfakes for undergraduate project. With the GANs updating everytime they lose does this refer to the deeplearning?

[@subodhi6]

Thank you..!

[@debayanguha3026]

thank you ,it's great ...!

[@apdy1095]

can someone tell me wht the core idea behind DDQN and GAN is same

[@heidikeller50]

Super- thank you :)

[@BintAlAbla1999]

Great video, very well done, thank you. I can see it can generate amazing imagery etc.. Allow me to ask a dumb question. What is the point of GANS? How does it enhance learning, for example? I just don't get 'the point'.

[@Behdad47]

Have you found your answer yet?

[@Has_Le_India13]

if we are giving the discriminator a domain for learning shapes of flower isnt is supervised learning how it is unsupervised since we are providing a domain to learn

[@alaad1009]

Excellent video

[@abdurrouf4159]

Well explained.

[@fundatamdogan]

I loved the lesson.But GANs more :)

[@user-uw1bb6rr8i]

Hey there, I am writing my bachelor thesis about how safe facial recognition authenticators will be with improving AI image creation. Would you say that GANs can oppose a risk to facial recognition authenticators? Thank you

[@EmpoweredWithZarathos2314]

Loved it😅

[@Callmejz.ai01]

if this is unsupervised, how does the discriminator "know better be able to tell where we have a fake sample coming in"? thank you for your theory, and the flower example! #creatoreconomy

[@AixinJiangIvy]

It‘s helpful. Finally know what GANs are, appreciate it.

[@Krunkbitmos]

the discrimator is trained a normal way with real flower pictures? how is the generator trained to make the first flower? like how does it know to output certain data in certain size and colors etc? i understand how it can update if wrong but how is the generator actually generating?

[@basedmatt]

Could somebody explain to me the difference between a GAN and Zero-Shot Learning?

[@blumehao]

you use right hand?

[@uurv]

is this possible to make a one image into different poses, variations. Can anyone reply to this image

[@hassanbinali1999]

Yes udaya it is possible. We call this method "data augmentation". You can find a lot of techniques on internet related to this.

[@MdAbdullah-gn6uj]

Nice video

[@jasonchen7758]

He is either a lefty that can write mirror image sentences from right to left in real time, or the video was post processed?

[@zlygerda]

Flipped

[@prateekkatiyar9532]

Great

[@golamrob]

excellent

[@Zackemcee1]

Is this what Nvidia is using for its new frame generation technique in the RTX 40 series? I'm just guessing before checking the internet

[@keshavmiglani2697]

Did DALL-E 2 use GAN?

[@storytimewithme2]

why don't you have a link to the CNN video that he mentions?

[@MdAbdullah-gn6uj]

Nice

[@MdAbdullah-gn6uj]

😊Nice

[@leif1075]

Didn't most everyone else think that is not what zeromsum game meant..inthoight if there is an advantage for one player that would not be a zero sum game..

[@Evokus]

Are we just going to ignore the fact that he's writing backwards??? That thing is skill man

[@uday3350]

Relax, he would have flipped the video left to right so that you don't see the text backwards.

[@tudorrad5933]

I literally spent the entire video not listening to him and asking myself what wizardry he uses to write mirrored.

[@Billy-sm3uu]

he wrote with his right hand then mirrored the video

[@drakefruit]

how do you write backwards so well lol

[@DavOlek_ua]

picture is mirrored? my brain is glitching and I don't know why lol

[@IBMTechnology]

Hey there! We shared some behind the scenes of our videos on the Community page, check it out here 👉 ibm.co/3dLyfaN 😉

[@DavOlek_ua]

@@IBMTechnology haha I knew it is exactly like that!)

[@saifshaikh8679]

Are Generators used for creating deep fakes?

[@techwithbube]

First to comment .

[@java2379]

I don't get that the discriminator should be updated if the generator succeeds. The image was 'fake' ( i would say synthesized ) and the whole point of the game beeing to teach the generator how to synthesize image that are as far as possible close to the 'real data' dataset. There is no failure per say. It all depends on what you means by fake: 1- Fake means even if its a realistic flower but does not belong to the 'real' dataset it a fake. 2- Fake means its not a flower ,its a car , or garbage so the discriminator is unhappy of the generator's job. You seem to define fake as per definition 1 ; in this case , you can directly compare image pixels by pixels and calculate euclidian distance for the error to backpropagate on the generator, you don't need a neural network for the discriminator , do you? So i think the correct definition is 2. Hence the discriminator never has to learn from the generator. >> I know you work for IBM , so its likely that i missed a point , kindly let met know 🙂

[@THEMATT222]

Noice 👍 Doice 👍 Ice 👍

[@erikschiegg68]

Gimme Ampere 100 Now! (GAN) Just for StyleGAN3, please, sir.

[@sc1ss0r1ng]

no, you give me 100 amperes now and also 1500 volt, madam. I will not ask twice, hand it over, or you will be shocked, by the consequences.

[@RuiMartins]

I hope the host understands that he could write normally, instead of reflected, since he just needs to mirror the video in the end and everything would be correct from the viewers view.

[@IshanJawade]

How can he write upside down

[@mewtu5817]

A gan is a speedcube

[@--Dipanshu--]

how is he writing backwards?

[@aryanarya72]

He's not writing backwards. It appears as if he is. He is writing normally like you would on a board or a notebook.

[@Steppinonshii]

what type of magis is this . he is writing backwards

[@IBMTechnology]

See ibm.biz/write-backwards for the backstory

[@Steppinonshii]

@@IBMTechnology omg 🤣🤦‍♂️

[@ShaliqAbu]

Avengers need you ,pls go back....

[@SheSweetLikSugarNSavage]

I've had a few supervisors that I'm sure were fake samples.😐

[@RudreshSisodiya]

If IBM don't have money for mirror marker, send me the bank details, I'll pay for it.

[@mariusulmer1932]

superb backwards writing

[@ChaojianZhang]

Feels like talking something but didn't tell much.

[@hi_dude_im_a_man]

No it’s a cubing company

[@zlygerda]

He's not really left handed, you know.

[@nikolakalev4914]

Are you really writing all of this backwards?

[@IBMTechnology]

Search on "lightboard videos".

[@jonobvious]

isn't it weird how all these glass whiteboard people are left handed. like usually about 10% of people are left handed but these guys I swear are like 90%, weird

[@albertxcastro]

Isn't the video mirrored horizontally? Otherwise I can't explain why we can see in the right direction what he's writing

[@chintanshah6234]

They're right handed..it is horizontally mirrored

[@petteruvdal3353]

Yo he writing backwards

[@IBMTechnology]

See ibm.biz/write-backwards

[@thechoosen4240]

Good job bro, JESUS IS COMING BACK VERY SOON; WATCH AND PREPARE

[@jotatd4038]

Bro just kept talking and said nothing

[@MdAbdullah-gn6uj]

Nice