# Google AI Simulates Evolution On A Computer! 🦖

## Метаданные

- **Канал:** Two Minute Papers
- **YouTube:** https://www.youtube.com/watch?v=a0ubtHxj1UA
- **Дата:** 25.06.2022
- **Длительность:** 8:40
- **Просмотры:** 216,674

## Описание

❤️ Check out Weights & Biases and sign up for a free demo here: https://wandb.com/papers 
❤️ Their mentioned post is available here (Thank you Soumik Rakshit!): https://wandb.me/modern-evolution

📝 The paper "Modern Evolution Strategies for Creativity: Fitting Concrete Images and Abstract Concepts" is available here:
https://es-clip.github.io/

🧑‍🎨 My previous genetic algorithm implementation for the Mona Lisa problem (+ some explanation in the video below):
https://users.cg.tuwien.ac.at/zsolnai/gfx/mona_lisa_parallel_genetic_algorithm/

❤️ Watch these videos in early access on our Patreon page or join us here on YouTube: 
- https://www.patreon.com/TwoMinutePapers
- https://www.youtube.com/channel/UCbfYPyITQ-7l4upoX8nvctg/join

🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
Aleksandr Mashrabov, Alex Balfanz, Alex Haro, Andrew Melnychuk, Benji Rabhan, Bryan Learn, B Shang, Christian Ahlin, Eric Martel, Geronimo Moralez, Gordon Child, Ivo Galic, Jace O'Brien, Jack Lukic, Javier Bustamante, John Le, Jonas, Jonathan, Kenneth Davis, Klaus Busse, Kyle Davis, Lorin Atzberger, Lukas Biewald, Matthew Allen Fisher, Michael Albrecht, Michael Tedder, Nevin Spoljaric, Nikhil Velpanur, Owen Campbell-Moore, Owen Skarpness, Rajarshi Nigam, Ramsey Elbasheer, Steef, Taras Bobrovytsky, Ted Johnson, Thomas Krcmar, Timothy Sum Hon Mun, Torsten Reil, Tybie Fitzhugh, Ueli Gallizzi.
If you wish to appear here or pick up other perks, click here: https://www.patreon.com/TwoMinutePapers

Thumbnail image: DALL-E 2.
Thumbnail background design: Felícia Zsolnai-Fehér - http://felicia.hu

Károly Zsolnai-Fehér's links:
Instagram: https://www.instagram.com/twominutepapers/
Twitter: https://twitter.com/twominutepapers
Web: https://cg.tuwien.ac.at/~zsolnai/

## Содержание

### [0:00](https://www.youtube.com/watch?v=a0ubtHxj1UA) Segment 1 (00:00 - 05:00)

Dear Fellow Scholars, this is Two Minute  Papers with Dr. Károly Zsolnai-Fehér. Today is going to be all  about simulating evolution   on our computers. This can build virtual cars,  virtual creatures, even paint the Mona Lisa!    These all sound amazing, but,  evolution on a computer? How? A few years ago, a really fun online app surfaced  that used a genetic algorithm to evolve the   morphology of a simple 2D car with the goal of  having it roll as far away from a starting point   as possible. A genetic algorithm? What is  that? Well, it is a super simple technique   where we start out from a set of random solutions,  and likely find out that none of them really work   well. However, we start combining and mutating,  or in other words, changing small parts of this   solution, until, oh yes! Something starts to  move. Now as soon at least one wheel is placed   correctly, the algorithm will recognize  that this one rolled so much further,   and keep the genes of this solution in the  pool to breed new, similar solutions from it. A similar concept can also be used to design  the optimal morphology of a virtual creature   to make it skip forward faster. And,  look! That is so cool - over time,   the technique learns to more efficiently  navigate a flat terrain by redesigning its legs   that are now reminiscent of small springs  and uses them to skip its way forward.    And here comes something even cooler - if we  change the terrain, the design of an effective   agent also changes accordingly, and the super  interesting part here is that it came up with   an asymmetric design that is able to climb  stairs and travel uphill efficiently. Loving it! And, in this new paper, scientists at Google Brain  tried to turbocharge an evolutionary algorithm   to be able to take a bunch of transparent  triangles, and get it to reorganize them so they   will paint the Mona Lisa for us. Or any image in  particular. The technique they are using is called   evolution strategies, and they say that it is  much faster at this than previous techniques. Well, I will believe it when I see it. Hold  on to your papers, and let’s see together.    Here is a basic evolutionary algorithm after 10  thousand steps. Well, it is getting there, but   it’s not that great. And, let’s see how well their  new method does with the same amount of steps.    Wow. My goodness! That is so much better. In fact,  their 10 thousand steps is close to equivalent   to half a million steps with a previous algorithm. And we are not done yet. Not even close. This  paper has two more amazing surprises. Surprise   number one. We can even combine it with OpenAI’s  technique called CLIP, this learns about pairs   of images and their captions describing these  images, which is an important part of DALL-E 2,   their amazing image generator AI. This  could take completely outlandish concepts   and create beautiful, photorealistic images out  of it that are often as good as we would expect   from a good human artist. Scholars holding  on to their papers, a cyberfrog, you name it. So, get this, similarly to that, we can now  even write a piece of text, and the evolutionary   triangle builder will try to create an image  that fits this text description. With that,   we can request a triangle version of a  self-portrait, a human, and look at how similar   they are. Food for thought! But it can also try to  draw Walt Disney World, that is remarkable. Look   at how beautifully it boils it down to its essence  with as few as 200, or even just 25 triangles.    Loving it. Also, drawing the Google  headquarters? No problem at all. And here comes surprise number  two. The authors claim that it is   even faster than a differentiable renderer.   These are really powerful optimization techniques   that can even grow this beautiful statue out of  nothing. And do all that in just a few steps. So,

### [5:00](https://www.youtube.com/watch?v=a0ubtHxj1UA&t=300s) Segment 2 (05:00 - 08:00)

claiming that it is even faster than  that? Well, that is very ambitious.    Let’s have a look. And…oh yes, as expected the  differentiable technique creates a beautiful image   very quickly. Now, try to beat that! Wow…the new  method converges to the final image even quicker.    That speed is simply stunning. Interestingly, they  also have different styles. The differentiable   renderer introduces textures that are not  present in the original image, while the new   technique uses large triangles to create a smooth  approximation of the background and the hair, and   uses smaller ones to get a better approximation  of the intricate details of the face. Loving it. Now, let’s ramp up the challenge, and test  these a little more and add some prompts. And,   oh yes, this is why the differentiable  technique flounders on the prompts. Look.    It almost seems to try to do everything  all at once. While, the new technique   starts out from a fair approximation, and  converges to a great solution super quickly. I hope that this new take on evolution  strategies will be plugged in to applications   where differentiable techniques do well,  and perhaps do even better. That would be   absolutely amazing because these are a treasure  trove of science-fiction like applications.    For instance, we would be able to  almost instantly solve this billiard   game, where we would like to hit the white  ball with just the right amount of force   and from the right direction, such that the  blue ball ends up close to the black spot. Or,   simulate ink with a checkerboard pattern, and  exert just the appropriate forces so that it   forms exactly the Yin-Yang symbol shortly after. Or, here comes a previous, science-fiction like   example. This previous differentiable technique  adds carefully crafted ripples to the water,   to make sure that it ends up in a state that  distorts the image of the squirrel in a way that   a powerful and well-known neural network sees  it not as a squirrel, but as a goldfish. Wow. And, if this new evolutionary  technique could do all of these tasks,   but better and faster? Sign me up  right now. What a time to be alive! Thanks for watching and for your generous  support, and I'll see you next time!

---
*Источник: https://ekstraktznaniy.ru/video/13526*