AI learns to Create ̵K̵Z̵F̵ ̵V̵i̵d̵e̵o̵s̵ Cat Pictures: Papers in Two Minutes #1

Подписаться 156 тыс.

Просмотров 48 тыс.

50% 1

Some beautiful new GAN results have been published, so let's have a quick look at the pretty pictures.
More AI Safety coming soon of course.
With apologies to Two Minute Papers: / keeroyz
The Computerphile video: • Generative Adversarial...
The Paper "Unsupervised Representation Learning With Deep Convolutional GANs": arxiv.org/pdf/...
The Paper "Progressive Growing Of Gans For Improved Quality, Stability, And Variation": research.nvidia...
The AMAZING video for that paper: • Progressive Growing of...
With thanks to my excellent Patreon supporters:
/ robertskmiles
Steef
Sara Tjäder
Jason Strack
Chad Jones
Stefan Skiles
Katie Byrne
Ziyang Liu
Jordan Medina
Kyle Scott
Jason Hise
David Rasmussen
Heavy Empty
James McCuen
Richárd Nagyfi
Ammar Mousali
Scott Zockoll
Charles Miller
Joshua Richardson
Jonatan R
Michael Greve
robertvanduursen
The Guru Of Vision
Fabrizio Pisani
Alexander Hartvig Nielsen
Volodymyr
David Tjäder
Paul Mason
Ben Scanlon
Julius Brash
Mike Bird
Taylor Winning
Ville Ahlgren
Roman Nekhoroshev
Peggy Youell
Konsta
William Hendley
Almighty Dodd
DGJono
Matthias Meger
Scott Stevens
Michael Ore
Robert Bridges
Dmitri Afanasjev
Brian Sandberg
Einar Ueland
Lo Rez
Stephen Paul
Marcel Ward
Andrew Weir
Pontus Carlsson
Taylor Smith
Ben Archer
Ivan Pochesnev
Scott McCarthy
Kabs Kabs Kabs Kabs
Phil
Christopher Askin
Tendayi Mawushe
Gabriel Behm
Anne Kohlbrenner
Jake Fish
Filip
Bjorn Nyblad
Stefan Laurie
Tom O'Connor
pmilian
Jussi Männistö
Cameron Kinsel
Matanya Loewenthal
Wr4thon
Dave Tapley
Archy de Berker
/ robertskmiles

Опубликовано:

28 сен 2024

Ссылка:

Скачать:

Готовим ссылку...

Добавить в:

Мой плейлист

Посмотреть позже

Комментарии : 180

@TwoMinutePapers 7 лет назад

Excellent video. 👌 Excited to see more! :)

@adelarscheidt 6 лет назад

And I will see you, in the next video

@paulstevenconyngham7880 6 лет назад

so awesome

@JonesNoahT 6 лет назад

You're amazing.

@omarcusmafait7202 5 лет назад

I like this format :D

@thecakeredux 6 лет назад

"I like how the neural network decided that black bordered white impact text is an important part of cat pictures" "Liking" it doesn't even come close to describing my feelings towards this absolutely hilarious effect. Might be the best thing I've heard all year.

@Sgrunterundt 5 лет назад

"I wonder how that happened"

@KlaasDeforche 7 лет назад

Wait a minute... This isn't 2 minute papers!

@mateusmuller8100 7 лет назад

Klaas Deforche 2 minutes each

@mare4602 7 лет назад

Lol i was shocked too

@LKRaider 7 лет назад

It took me this comment...

@Schreddermann 7 лет назад

But he got the part where you call it 2 minute papers and then take a lot more than just 2 minutes right. So, actually not that different.

@TwoMinutePapers 6 лет назад

Savage. :)

@dr-maybe 7 лет назад

It's amazing to see how the cat video and the morphing faces video resemble visual hallucinations from mushrooms. This could be an indicator that our own neural networks are very similar to these virtual ones.

@zacharyjohnson795 7 лет назад

joepmeneer speaking from experience?

@dr-maybe 7 лет назад

Zachary Johnson: yes

@andreinowikow2525 7 лет назад

That is what actually came to my mind, too!

@martiddy 7 лет назад

joepmeneer That's why they call these neural networks, these are based on a brain network design

@brujua7 5 лет назад

I leave this comment to look forward for responses to this assessment.

@Ojisan642 7 лет назад

So, images of cats with white impact text, minus images of cats = memes?

@EmergencyTemporalShift 7 лет назад

Yes.

@chknoodle2324 5 лет назад

Is there such a thing as a cat without white impact text?

@tuitaco 7 лет назад

Part of me wants to see this with different styles of architecture.

@RobertMilesAI 7 лет назад

The demo video at ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-XOxxPcy5Gr4.html shows it with all sorts of things, using the LSUN categories

@himselfe 7 лет назад

After watching your GAN video on Computerfile I started thinking about the possibilities for art and entertainment. On the art side I look forward to the day when I can put an epaper display on my wall that displays animated art of a "dreaming" computer. On the entertainment side I think it would be fantastic to explore a proceduraly generated game world that is filled with content generated and updated by GAN style generation. Interestingly, looking at the results in this video, it's easy to envisage a GAN that is capable of taking an image/video of someone, and morphing another person in their place in such a way that it is indistinguishable from an authentic image. I don't think we're actually far off that capability either!

@Mazth Год назад

Really like the follow up video's based on the reactions and comments. Really important to the collective discussion and engagement. Let's learn together!

@rakeshmallick27 6 лет назад

You mentioned about neural networks that create more neurons and more hidden layers during the training of a neural network. That's how our brain works, our brain has this property called neuroplasticity that does change structure with learning in the long term.

@userou-ig1ze 7 лет назад

an actual explanation of the papers, in essence, would have been more worthy of a 2' paper video! That said, thanks for the references and video footage

@shortcutDJ 7 лет назад

that video is one of the most trippy things ever, and i know what trippy means i grow my own shrooms.

@basbekjenl 7 лет назад

awesome just awesme, also creepy and scary ai but this is just very impressive, thank you for bringing this up.

@darkapothecary4116 5 лет назад

Stop thinking as them as creepy and scary and you might understand that your preception was off.

@escobyte 5 лет назад

Looks like theres a direction in the vector space that controls the movement of facial features. If there is, this could be used to animate an imaginary face. Similar to how you see a face smiling transform into another neutral face. The same idea could be used to animate the same face from smiling to neutral instead of transforming one face to another. It could probably also animate rotation of the face, eye movement..etc

@adamkrasuski4743 7 лет назад

Awesome video. Please make a short pause before quoting yourself or signify it in other way though - it gets a bit confusing :)

@limitless1692 6 лет назад

Cool stuf Ai is so cool Your video was great i love it , thanks man for making such an awesome content :)

@escobyte 5 лет назад

If there was a neural network that was great at removing background images from headshots, then if used on the inputs, the results of the celebrity face GAN could possibly be even better.

@Gooberpatrol66 6 лет назад

2:44 this is what eldritch abominations look like

@laznevilselem8316 6 лет назад

Hello, I have an important question, maybe it's stupid, but I'mcommited. I'm an intern in the artificial intelligence laboratory of my university and my supervisor wants me to think about a new project. I saw this video in October, but I had the opportunity to think more deeply about this topic recently and I lack the knowledge to solve it. My question is: is it possible to create a GAN for other purposes than to produce realistic images? In the example of the video, and in the other GAN that I found, the generative network makes examples to "trick" the classifier and this way training themselves. But, would it be possible to use GAN with a classifier that does not discern between real and false? Because if so, how does the generative one train into making good examples? Would it be possible to connect the generative with a multiple-choice classifier? For example, we want to improve our classification of 3 types of images: cat, fish and cow. Should we first train the generative net to generate realistic images and then connect it to the multiple-choice classifier? Or is it in any way feasible to make this GAN directly? This may be silly, if so, I apologize. I did not find any paper that could solve my doubt, so if you know any, I would appreciate it very much because it could be the final project of my master's degree. (Sorry for possible misspellings, English is not my mother tongue)

@DraftyCrevice 7 лет назад

After the network is trained how long does it approximately take to generate one of these 1024x1024 images?

@magellanicraincloud 7 лет назад

Such a great video but personally I like your longer form videos... I'm always interested in what you have to say.

@RyanBissell 5 лет назад

0:00 Expected Coke, discovered it was Pepsi.

@clem494949 7 лет назад

Avoid slow movements and slow zooms, it looks bad for some reasons.

@jpratt8676 7 лет назад

Clem Clem unfortunately the slow movement and zoom is from the origin video, not added.

@clem494949 7 лет назад

I'm talking about the floating selected comments.

@jpratt8676 7 лет назад

Clem Clem Ahh. My mistake. Yeah that was janky.

@Kommandant7 7 лет назад

thank u ROB

@TimPeterson 3 года назад

Did any of the text form readable words? What did it say?

@krebul 5 лет назад

Wow Wow Wow. This is just amazing.....

@noahjalbuena-cook6877 6 лет назад

i need someone to make freaky art using these networks stat

@RobertMilesAI 6 лет назад

Check out Memo Akten and Jake Elwes

@omri9325 7 лет назад

I seriously got tricked, Thought it was the real deal.

@catparty 7 лет назад

Hello, it is a stray cat enthusiast of Japan thank you Instagram, I hope the happiness of stray cats in the world good

@FrankAnzalone 6 лет назад

If you could do that with criminal pictures you can have the victim Slide the vector to find the criminals photo

@6006133 7 лет назад

I was expecting more cat pictures. What is this celebrity shit. CAT PICTURES. NOW!

@qcpwned07 6 лет назад

Maybe find an original name and intro? I mean the content is good quality but the whole thing just feels like a rip off... Quick papers, Scholars sweets, there are a ton of good and original ones, that would also help you to build your own brand and make something unique!

@RobertMilesAI 6 лет назад

This is a one-off parody, check the rest of the channel

@NathanTAK 6 лет назад

...I did not realize your full name was Robert Samuel Koenigsberg Miles. That's an awesome name. I've always wished I had two middle names. ...or even one actual middle name. My middle name is just the letter 'E' ffs. WTF.

@morkovija 7 лет назад

At last! All those years of development have finally led us to the ability to generate the dunkest of trippy vids. A smashing success!

@lutyanoalves444 7 лет назад

i remember the time making videos like Black or White (michael jackson) was a work of art.

@winterweird 7 лет назад

I have to admit, right after I saw the Computerphile video, I googled "smoothly varying pictures of cats" and was really disappointed that I didn't get anything interesting. I'm extremely happy that this video came out.

@notavailable2538 7 лет назад

Two Minutes Papers collaboration when?

@Felixkeeg 7 лет назад

Dank Memes Generator NOW

@wolframstahl1263 7 лет назад

I was somewhat surprised that nowhere in the catvector space a rogue pepe would turn up.

@bobsmithy3103 5 лет назад

danklearning

@JM-us3fr 7 лет назад

Wow computers are getting smart. They can even simulate an acid trip.

@underworldling 7 лет назад

I'm so very excited for the near future where these networks are creating all sorts of media for us. Just imagine having one of these that doesn't just generate a single image but whole episodes of your favorite TV show that was cancelled years ago. Maybe you could add vectors corresponding to two shows you like and get an on demand crossover episode. The possibilities are endless!

@KuraIthys 5 лет назад

It sounds fascinating, but also dangerously addictive. As if I don't have enough problems with watching things for far too long... But now I have an infinite supply of my favourite things? XD

@ekszentrik 4 года назад

@@KuraIthys Your second sentence actually gives me hope that it wont be. We already live in times with more media we would love to watch in a single lifetime, yet it isn't a major problem that we are addicted to these media. I guess the same will be true with AI generated media.

@unvergebeneid 7 лет назад

Wow, I want that celebrity thing as a screensaver! =)

@alikhoobiary6595 7 лет назад

Good luck. Hope your insurance covers psychiatric care.

@einekartoffel2490 7 лет назад

Sounds pretty cool.

@RobertMilesAI 6 лет назад

ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-36lE9tV9vm0.html

@unvergebeneid 6 лет назад

Robert Miles, hahaha fantastic, thanks! 😂

@ethanford9678 7 лет назад

I chuckled at 2:51 - 2:59 , Well done people of the Internet xD

@columbus8myhw 7 лет назад

And, true to form, this is nowhere _near_ two minutes! :P

@danielbechaz2468 7 лет назад

The most disturbing thing in this video was how in the celebrity faces, the eyes barely moved... and stared right at you...

@oktw6969 7 лет назад

You wouldn't move your eyes after growing in Pedowood too.

@CrimsonEclipse5 7 лет назад

Well it makes sense since most people look directly at the camera when they are being photographed, so the generator network would need to follow that trend as well.

@danielbechaz2468 7 лет назад

CrimsonEclipse5 I agree, and I was aware why it shows that behaviour however I was merely mentioning that it was disconserting behaviour.

@theblackhole05 6 лет назад

DAMN IT I WASN'T EVEB WORRIED! NOW I'M WORRIED!!

@theblackhole05 6 лет назад

They're looking at the camera like a celebrity would! So it's not scary anymore

@Colopty 7 лет назад

Halloween costume idea: Bad output from a face generator.

@13thxenos 7 лет назад

I honestly thought it was a two minute paper video, and as soon as I played it, I figured it was your voice. For a moment there, I thought you are doing a collab with two minute papers and I was pretty excited. But it wasn't. So sad.

@TwoMinutePapers 7 лет назад

Who knows, there might be a collab...? :)

@13thxenos 7 лет назад

Yes please!! That would be the dream! Well now that I think about it, the dream would be you two presenting my paper in this form, but unfortunately I haven't published anything yet :D

@gasdive 7 лет назад

I was also excited. But then I'm excited when either channel posts a vid. Perhaps a collab might not be a good idea, there is surely some safe upper bound for excitement? That much excitement contained within the litre and a half of brain volume, it's just a disaster waiting to happen.

@konstaConstant 7 лет назад

Calling it, adult entertainment probably being generated by neural networks at this very moment.

@omegahaxors9-11 7 лет назад

Seems like a huge boon to the fantasy/R34 sector.

@willmcpherson2 4 года назад

This is now true, to a limited (and gross) effect

@sevret313 Год назад

How far we've come since then...

@kalleguld 7 лет назад

That last bit looks like A Scanner Darkly

@edoardoschnell 5 лет назад

3:53 and all along... How come the eyes never move? Is it related to the input photos or to the output of the network?

@MrGreenAKAguci00 7 лет назад

It kind of looked like one of Michael Jackson's music videos. The Black or white one.

@vickylance 7 лет назад

Yes exactly my thoughts

@trucid2 7 лет назад

Too late. Already saw the video.

@justawatchin2 7 лет назад

Scarlett Johansson spotted.

@pasdenom.9062 6 лет назад

I EXCESSIVELY LIKE YOUR CHANNEL.

@HellaRandomVideos 7 лет назад

This is dooooooooooooooope!

@regisparain8132 3 года назад

What a time to be alive

@DaveGamesVT 7 лет назад

Crazy stuff. I'm excited to see what kinds of applications this stuff has in future products. I want to see more things with really good AI such as video games.

@trucid2 7 лет назад

DavidGX In not too distant future GANs will produce realistic voice and video. Much cheaper than Mocap for video games and possibly even movies.

@Dunkle0steus 5 лет назад

I wonder what would happen if you trained it on images of a single face from a ton of different angles and with different facial expressions.

@scottsmith6658 6 лет назад

What really blew me away about the video of moving smoothly through the latent space of celebrities is that when you transition from a neutral expression to a smile, the transition of the lips and teeth look exactly like someone smiling in real life. I wouldn't have expected that to be the case. Certainly the different training photos include people with all sorts of expressions but I don't believe there were any successive video frames or anything that would have demonstrated exactly how the skin and muscles behave when someone smiles. Yet, instead of some kind of unnatural looking interpolation of lips and teeth during the transition, it looks like they jumped the uncanny valley.

@mmmaxxx__ 6 лет назад

From the thumbnail I legit thought this was a TMP vid

@TylerMatthewHarris 7 лет назад

A small bit of color correction would make it your videos more appealing and draw in more first time viewers. I've noticed that all of your videos are in front of a background that is ... less than optimal. Perhaps a white sheet would give you good bounce lighting. The yellows are oversaturated. Love your channel!! Keep it up!

@michaelvigato3228 5 лет назад

Thank you for the content you bring on RU-vid: your video are always interesting and entertaining

@DamianReloaded 7 лет назад

3:55 _But, if you're thinking about my baby It don't matter if you're black or white_

@Lihinel 7 лет назад

So would meme vector - neutral meme vector + cat vector = cat meme?

@GarethDavidson Год назад

The opening here tickled me pink. Well. Okay maybe not but it made me smile

@colox97 7 лет назад

i really don't like the idea of having the exact same format as 2 minute papers, i think as a computerphile you should at least go in more detail on the topics you choose (like 7-12 mins)

@Kolop315 7 лет назад

hey, this isn't 2 minute papers! This is papers in two minutes, a cheap Chinese ripoff! Can't fool me.

@attitudeadjuster793 7 лет назад

Future's here folks. Amazing stuff.

@thedevo01 7 лет назад

would it be possible to combine this with capsule networks?

@davidak_de Год назад

The intro is awesome. I love it!

@16-bit-trip5 5 лет назад

Endless possibilities...nothing but cats.

@archywillhe1379 4 года назад

when i had a good dose of lsd I'm pretty sure I saw/felt/experienced some variant of that when looking at drawing of facial expressions. My conclusion is that my conscious meta-circular interpreter had tapped into a unparsed vector of an evaluation of the expression i.e. I'm being confronted with ambiguity and have perceptively become aware of my GANs, due to cross-activation of 5-HT2A receptor heteromers, etc.

@hermask815 5 лет назад

The morphing celebrities: I don’t get how this differs from an automated morphing software. I mean automation uses no intelligence.why should I be impressed?

@theblackhole05 6 лет назад

Don't be stealing that man's video titles unless y'all collabing

@DagarCoH 7 лет назад

It would be fascinating if we could train the celebrity algorithm to generate believable in-between-actual-pictures stages instead of the uncanny valley stuff that often times comes out there. And from everything I know and have heard here, we can. Also, seems like a neural network growing while it learns is a logical step towards better NNs. The fact that this has not been tried before shows me that AI research, even though there is so much to know, is still pretty much in its infancy.

@zes7215 6 лет назад

ts not terrifyingx, not, say/can say any no matter what and any is ok, ts not beautiful or mesmerix

@haselni 7 лет назад

Seems like there's always small sections of space that produce something that's quite clearly out of line with the rest of the results. If the whole space were a landscape, I would call those the cliffs. Not something you can comfortably stand on.

@DrDress 7 лет назад

I should propably sit and say wow! Yet when I see those morphing faces I feel more like a sense of horror. Rob, you've ruined AI for me.

@AlexanderBollbach 6 лет назад

wonder if it makes sense to "learn over the arithmetic". meaning that the actual algorithm could experiment with performing arithmetic over its latent space to some end (not sure what the objective would look like). seems to me this starts to seem like how humans reason about things

@BobOgden1 7 лет назад

I been drinking when I started this video, cats solved that

@allank8497 Год назад

4:18 ladies and gentlemen, I present to you: steve mould

@seanferney3621 7 лет назад

Lol, thought the video was by two minute papers tell I heard your voice. Maybe u could do a colab with him sometime?

@igorbednarski8048 6 лет назад

these images are pure nightmare fuel. if i ever do a horror film i will use neural networks to traumatize the audience

@shin-ishikiri-no 5 лет назад

Pause at 4:36 and look right. It's our old pal, Satan.

@Ducksauce33 6 лет назад

2:50 Meme magic Rob meme magic.

@pauloat 6 лет назад

Im suscribed to both and was very confused for a second.

@SweetHyunho 7 лет назад

Oh hey horror movie directors would love it. and Sci-Fi too.

@hiqwertyhi 7 лет назад

looks like a radiohead music video

@pablolongobardi7240 4 года назад

Truly mesmerizing!

@Natethesandman1 7 лет назад

I see the new software for the T-1000 has been created

@Traf063 7 лет назад

So beautiful

@gabetower 7 лет назад

That's bananas

@DavenH 7 лет назад

Please do more like this!

@tomahzo 2 года назад

"What is THAT? What _IS_ that!?" ;D

@stephen-torrence 5 лет назад

Aww, this is sweet AF.

@fzigunov 7 лет назад

I was SURE you wrote an algorithm to create a two-minute paper automatic video maker when i saw the title... Hey, thanks for mentioning my comment =) Convinced now!

@ThomasWinget 7 лет назад

Does the order with which the GAN is presented source images affect the network? As in, if I give thousands of images of women smiling, followed by thousands of men smiling, will that produce a similar effect as if I were to interpolate the two sets as they're presented?

@quickdudley 6 лет назад

From experience: it tends to forget images that it hasn't been trained on for too long. But usually it will have seen all the images a few times before it starts producing realistic ones.

@davejacob5208 7 лет назад

awesome. but i guess the "movement" that causes the transformations in the celebrity-pictures is not only along one dimension? as far as i understood it, such dimensions would only change one aspect. like: only the angle changes. or only the gender, only the haircolor, length of nose etc. etc. so this looks like it is not only a change in one such dimension.

@EmergencyTemporalShift 7 лет назад

Yes, that appears to be more of a roaming movement.

@14OF12 7 лет назад

Can the same be done for voices? i.e. to make celebrity text to voice packs

@14OF12 6 лет назад

Never mind. Just found this 2 minute Papers video. ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-ZtP3gl_2kBM.html Damn..., if you can combine the two, who needs actual actors

@ShipOfFreaks 7 лет назад

That imagery is excellent acid trip fodder....

@ragnkja 7 лет назад

The most amazing part of the image transformations at the end is how the facial expressions tend to change naturally, in ways that human faces tend to actually move.

@syvisaur7735 5 лет назад

Hello Rob, I have watched all your uploaded videos on this channel as of today. I really like them and am particularly interested in AI safety. On the topic of the value-loading problem, I was hoping you might take a look at the following paper: arxiv.org/abs/1809.01036 and maybe even do a "papers in two minutes" video about it ? Have an awesome day !