This content was created by LARRUZ. This content was uploaded by LARRUZ. This is a YouTube video titled "New AI Beats DeepMind’s AlphaGo Variants 97% Of The Time!" (published 2023-08-21) featuring the 2 Minute Papers presenter (Dr. Károly Zsolnai-Fehér). The excerpt is an explainer-style presentation about adversarial attacks on modern AI, with concrete examples: adversarial agents that systematically exploit Go-playing neural nets (claiming an attack that defeats KataGo ~97% of games), single-pixel image attacks (horse→frog), and crafted-noise attacks that fool image classifiers (e.g., Van Gogh’s Starry Night). The speaker emphasizes the surprising vulnerabilities of powerful AI systems and the need for more research, briefly mentions LK99, and includes a sponsor plug from Weights & Biases. The transcript mixes Hungarian and English and appears to be a monologue with a short sponsor exchange.
Questions this video answers
97% of games.
2023-08-21.
Chapters
Köszönjük a feliratkozatokat! Ez volt 2 Minute Papers között Dr. Károly Zsolnai-Fehér. Well, of course, not with brute force, but with trickery. This is what we call an adversarial attack. And few know about it, but many modern AI techniques are quite vulnerable against them. This is the you shall not pass game, where the red agent is trying to hold back the blue character and not let it cross the line. Igen, itt látod, két egész AI-t kipróbálja, valahol a kérek védik, valahol a kérek meg tudja átjutni. Nincs semmi ilyesmi itt. Ez a számítókép, amely egyáltalán jól találkozik. Most nézzétek közel, mert itt jön a hackers adversarial agent. Yes, you are seeing correctly, this chap is doing nothing. Absolutely nothing. But it is doing nothing in a way that reprograms its opponent to make mistakes and behave close to a completely randomly acting agent. This paper was absolute insanity. A different adversarial attack paper showed an interesting case where we were able to take an image of a horse. The attacked neural network indeed recognized that this was a horse. However, when changing just one pixel, an otherwise quite competent AI now thinks that this is a frog. Note that this does not mean that we can just change any pixel anywhere. This is a sophisticated attack that knows who it is attacking, and how this one pixel difference will reprogram its brain. Similar attacks exist that do this with a little more flavor. This is a buss, and this is noise. And when we add this together, do we get buss plus noise? Nope, what we get is an ostrich. So does the AI think. Now, let's have a look at new papers with new attacks, first against AIs that can play Go. Now, wait a second, what is really new here? We have heard players reporting before that with AlphaZero, they have experienced little hiccups where the AI made suboptimal moves in an otherwise really well played game. Why is that news? Why write a paper about this? How is this one different? Well, what makes this attack interesting is that this is not a one-off fluke, this is an AI that finds systematic flaws in other neural network-based systems. This means that they are not only able to exploit their weaknesses and win against them, but they can do it on a consistent basis. In other words, they can win over and over again. In this case, hold on to your papers, because this paper describes an attack that is able to defeat Katago in 97% of the games. That number is unreal. Let me explain why. From what I have seen, Katago seems to be even stronger than AlphaZero and AlphaGoZero, DeepMind's legendary systems that are able to beat the best human players in the world. The authors note that it works on many AlphaGo-based systems, and it likely works against AlphaGo variants too. What makes it even more impressive is that this adversary was trained from scratch and did not use any human knowledge, and found this all out by itself! And finally, here is another attack against a competent image recognition AI, when showing it the Starry Night painting by Van Gogh, which it will, of course, recognize as shown by the red frame. However, as we start adding carefully crafted noise to it, nothing happens. But just wait a little more, as we continue the process, damn! Now this noise looks nothing like Starry Night, but not according to this AI. It would swear that this is exactly the Starry Night painting. And this works for other examples too. The goal of all this is to show you some of the weaknesses of recent AI systems. This is a really interesting new field where we have these powerful new tools, but this also means that they have their own limitations too, and sometimes it is not at all obvious what they are. More research is required. Now, one word about the LK99, the Room Temperature Superconductor Project. I see that many of you Fellow Scholars would love an episode on it, however, I do not have the required knowledge to comment on it. Obviously, it would be a very good thing for views, but that doesn't matter. What matters is that you Fellow Scholars get the quality of videos that you expect here.
That is what matters. This video has been supported by Weights Biases. Look at this, they have a great community forum that aims to make you the best machine learning engineer you can be. You see, I always get messages from you Fellow Scholars telling me that you have been inspired by the series, but don't really know where to start. And here it is. In this forum, you can share your projects, ask for advice, look for collaborators and more. Thanks for watching and for your generous.
