DeepMind's MU0, an AI trained on games like chess and Go, tackled video compression and beat the industry standard VP9 codec by about 4%. The agent learned to choose encoder settings that shrink file sizes while keeping visual quality high, showing that game-playing AI can solve real-world problems.
Dear fellow scholars, this is Two Minute Papers with Dr. Zsolnai. Finally today, DeepMind's amazing AI MU0 that plays chess and other games has now finally entered the real world and has learned to solve important real world problems. This is a reinforcement learning technique that works really well on games. Why? Well, in Chess Go and starcraft, the controls are clear. We use the mouse to move our units around or choose where to move our pieces. And the score is also quite clear. We get rewarded if we win the game. That is going to be our score. To say that these worked really well would be an understatement. DeepMind's Mew0 is one of the best in the world in Chess Go and Starcraft 2 as well. But one important question still remains. Of course they did not create this AI to play video games. They created it to be able to become a general purpose AI that can solve not just games, but many problems. The games are just used as an excellent testbed for this AI. So what else can it do? Well finally here it is. Hold on to your papers. Because scientists at DeepMind decided to start using their MU0AI to create a real solution to a very important problem, Video compression. And here comes the twist. They said lets imagine that video compression is a video game. Okay, that's crazy, but let's accept it for now. But then two questions. What are the controls and what is the score? How do we know if we want video compression? Well, the video game controller in our hand will be choosing the parameters of the video encoder for each frame. Okay, but there needs to be a score. So what is the score here? How do we win? Well, we win if we are able to choose the parameters such that the quality of the output video is as good as the previous compression algorithms. But the size of the video is smaller. The smaller the output video, the better. That is going to be our score. And it also uses self competition, which is now a popular concept in video game AIs. This means that the AI plays against previous versions of itself and we measure its improvement by it being able to defeat these previous versions. If it can do that reliably, we can conclude that yes, the AI is indeed improving. This concept works on boxing, playing catch, starcraft. And I wonder how this would work for video compression. Well, let's see, let's immediately drop it into deep waters. Yes, we are going to test this against a mature state of the art video compression algorithm that you are likely already using this very moment as you are watching this on YouTube. Well, good luck little AI but I'll be honest, there is not much hope here. These traditional video compression algorithms are a culmination of decades of ingenious human research. Can a newcomer AI beat that? I am not sure. And now hold on to your papers and let's see together. How did it go? So, a 4% difference. A learning based algorithm that is just 4% worse than decades of human innovation. That is great. But wait a second, it's actually not worse. Can it be that. Yes, it is not 4% worse. It is even 4% better. Holy mother of papers. That is absolutely incredible. Yes, this is the corner of the Internet where we get super excited by a 4% better solution and understand why that matters a great deal. Welcome to two Minute Papers. But wait, we are experienced fellow scholars over here. We know that it is easy to be better by 4% in size at the cost of decreased quality. But having the same quality and saving 4% is insanely difficult. So which one is it? Let's look together. I am flicking between the state of the art and the new technique. And yes, my goodness, the results really speak for themselves. So let's look a bit under the hood and see some more about the decisions the AI is making. Whoa, that is really cool. So what is this? Here we see the scores for the previous technique and the new AI. And here they appear to be making similar decisions on this cover song video. But the AI makes somewhat better decisions overall. That is very cool. But look at that. In the second half of this gaming video. MU0 makes vastly different and vastly better decisions. I love it. And to have a first crack at such a mature problem and manage to improve it immediately, that is almost completely unheard of. Yet they have done it with protein folding and now they seem to have.
Done it for video compression too.
Bravo, DeepMind. And note the meaning in the magnitude.
Of the difference here.
OpenAI's Dall E2 was this much better than Dall E1. That's not 4% better. If that was a percentage, this would be several hundred percent better. So why get so excited about 4%? Well, the key is that 3 to 4% more compression is incredible given how polished these state of the art techniques are. VP9 compressors are not some first crack at the problem. No, no. This is a mature field with decades of experience where every percent of improvement requires blood, papers and tears. And of course, lots of compute and memory. And this is just the first crack at the problem for DeepMind. And we get not 1%, but 4% essentially for free. That is absolutely amazing. My mind is blown by this result. Wow. And I also wanted to thank you for watching this video. I truly love talking about these amazing research papers and I am really honored to have so many of you fellow scholars who are here every episode enjoying these incredible works with me. It really means a lot. Every now and then I have to pinch myself to make sure that I really get to do this every day. Absolutely amazing. Thank you so much. So what do you think? What else could this be useful for? What do you expect to happen a couple more papers down the line? Please let me know in the comments below. I'd love to hear your thoughts. This episode has been supported by Cohere AI.
Cohere builds large language models and makes them available through an API so businesses can add advanced language understanding to their system or app quickly. With just one line of code, you can use your own data, whether it's text from customer service requests, legal contracts, or social media posts, to create your own custom models to understand text or even generate it. For instance, it can be used to automatically determine whether your messages are about your business hours, returns, or shipping. Or it can be used to generate a list of possible sentences you can use for your product descriptions.
Make sure to go to Cohere AI.
Papers or click the link in the video description and give it a try today.
It's super easy to use. Thanks for watching and for your generous support and I'll see you next time.
