YouTube video (Two Minute Papers, Dr. Zsolnai) reviewing DeepMind research on an AI that learns from human instruction–video pairs in randomized virtual environments. The host demonstrates the agent performing diverse tasks (finding a bathroom, putting grapes in a bucket, answering object-count/color questions, playing a drum), highlights evidence the agent learns general concepts (randomized layouts, memory of prior dialogue), reports rapid learning (noticeable improvement after ~12 minutes of human interaction) and steady linear performance growth, cites ~70% success on some tasks, and discusses sim-to-real implications (robot hand Rubik’s cube, self-driving car training).
Chapters
Dear fellow scholars, this is Two Minute Papers with Dr. Zsolnai. Today we are going to see what DeepMind's amazing AI can learn from humans doing silly things like playing the drums with a comb. Really? We have to do that? Well, okay, if you insist. There you go. Well, if it had seen enough of these instruction video pairs, hopefully it will be able to learn from it. And here comes the fundamental question we are here to answer. If this AI does something, is this imitation or is this learning? Imitation is not that impressive. But if it can learn general concepts from examples like this, now that would perhaps be a little spark of intelligence. That is what we are looking for. So how much did it learn? Well, let's see together. Nice. Look at that. If we ask it, it can now show us where a specific room is. And if we get a little lost, it even helps us with additional instructions to make sure we find our way into the bathroom. Now, note that the layout of this virtual playhouse is randomized each time we start a new interaction. Therefore, it cannot just imitate the direction people go when they say bathroom. It doesn't work because next time it will be elsewhere. It actually has to understand what a bathroom is. Good job, little AI. It also understands that we are holding a bucket and it should put the grapes in that. Okay, so it learned from instructions, and it can deal with instructions. But if it had learned general concepts, it should be able to do other things too. You know what? Let's ask some questions first, how many grapes are there in this scene? Two. Yes, that is correct. Good job. Now, are they the same color or. Or different? It says different. I like it. Because to answer this question, it has to remember what we just talked about. Very good. And now hold on to your papers and check this out. We ask what color the grapes are on the floor and what does it do? It understands what it needs to look for to gather the required information. Yes, it uses its little peepers and answers the question. A little virtual assistant AI I love it. But it isn't just an assistant. Here come two of my favorite tasks from the paper. One, if we dreamed up a catchy tune and we just need a bandmate to make it happen. Well, say no more. This little AI is right there to help. So cool. And it gets better. Get this too. It even knows about its limitations. For instance, it knows that it can't flip the green box over. This can only arise from knowledge and experimentation with the physics of this virtual world. This will be fantastic for testing video games. So, yes, the learning is indeed thorough. Now let's answer two more super important questions. One, how quickly does the AI learn? And is it as good as a human? So learning speed, let's look under the hood together and oh yes, yes, yes. I am super happy. You are asking why are you super happy? I am super happy because learning is already happening. By watching humans mingle for about 12 minutes. That is next to nothing amazing. And one more thing that makes me perhaps even happier and that is that I don't see a sudden spike in the growth. I see nice linear growth instead. Why is that important? Well, this means that there is a higher chance that that the algorithm is slowly learning general concepts and starts understanding new things that we might ask it. And if it were a small spike, there would be a higher chance that it had seen something similar to a human clearing a table in the training set and suddenly started copying it. But with this kind of growth, there is less of a chance that it is just copying the teacher. I love it. And yes, this is the corner of the Internet where we get unreasonably excited by a blue line. Welcome to two Minute Papers. Subscribe and hit the bell icon if you wish to see more amazing works like this. And if we look at the task of lifting a drum. Oh yes, also nice linear growth, although at a slower pace than the previous task. Now let's look at how well it does its job. This is TeamHuman and MIA is Team AI and it has an over 70% success rate. And remember, many of these tasks require not imitation, but learning. The AI sees some humans mingling with these scenes, but to be able to answer the questions is it needs to think beyond what it has seen. Once again, that might be perhaps a spark of intelligence. What a time to be alive. But actually, one more question. A virtual assistant is pretty cool, but why is this so important? What does this have to do with our lives? Well, have a look at how OpenAI trained the robot hand in a simulation to be able to rotate these Rubik cubes and then deployed this software onto a real robot hand. And look, it can use this simulation knowledge. And now it works in the real world too. But Sim 2 Real has relevance to self driving cars too. Look, Tesla is already working on creating virtual worlds and training their cars there. One of the advantages of that is that we can create really unlikely and potentially unsafe scenarios. But in these virtual worlds, the self driving AI can train itself safely and when they deploy them into the real world, they will have all this knowledge. Waymo is running similar experiments too. 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. Thank you so much. So I think this one also has great potential for a Simtoria situation, learning to help humans in a virtual world, and perhaps uploading the AI to who? A real robot assistant. 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 Lambda GPU Cloud. If you're looking for inexpensive cloud GPUs for AI, check out Lambda GPU Cloud. Get this they've recently launched an Nvidia RTX A6000 with 48 GB of memory. And hold on to your papers because Lambda GPU Cloud can cost less than half of AWS and Azure. Join researchers at organizations like Apple, MIT and Caltech in using Lambda Cloud instances, workstations or servers. Make sure to go to lambdalabs.compapers to sign up for one of their amazing GPU instances today. Our thanks to Lambda for their long standing support and for helping us make better videos for you. Thanks for watching and for your generous support and I'll see you next time.
