Jason Yosinski, co-founder of Uber AI Labs, explains how machine learning has exploded in the past decade due to more compute power and larger datasets. He walks through three examples: DeepMind's AlphaGo beating the world Go champion, his own 3-D-printed robot that learned to move nine times faster using neural networks, and transformer models that generate plausible text like "Ovid's Unicorn" from billions of web pages.
Chapters
Progress is kind of skyrocketing, and this is due to a couple factors. First, we have this, like, volume of computation that's available. It's much larger now than it used to be. There's also data, more and more data available, natural language processing and other domains. And this increase in compute and data kind of directly has led to better and more powerful models. But we found that basically faster computers and more data kind of directly translate into just better ML, better AI. Thanks for that intro, and can you guys hear me? Sounds like this is on. Very cool hairstyle. Oh, thank you. Yeah, I just woke up. So I don't actually lead Uber AI Labs, although I was one of the co-founders. Just want to make that straight. I'm a researcher at Uber, and I work on AI and machine learning. Who here does anything with machine learning at all? Okay, a little, a few people. What do you do with machine learning? So I have a study in medical technology. We use AI in order to predict epilepsy seizures. To predict what? Epilepsy seizures. OK, OK. Based on what type of data? Biosensors, basically. OK, OK. Also, there's like your personal data, like age, height. OK, got it, got it, got it. That's cool. That's cool. Anyone else want to give an example? Yeah, please. Okay, what, what application? Mostly social media. Okay, NLP for social media. Cool, cool. I think when I came here in the spring, maybe nobody used AI for anything, or maybe like 1 or 2 people said they used it for something. So I'm going to talk to you a bit today about some of my research and research from collaborators in machine learning. Broadly, I'm going to talk about ML, why it works well, like how it works, What we know about how it works and what we don't know. Let's just get started quickly. So if we look at, like, machine learning progress over the last 5 or 10 or more years, we've seen basically, like, progress is kind of skyrocketing, and this is due to a couple factors. First, we have this, like, volume of computation that's available. It's much larger now than it used to be. So in log space, you know, in linear space, this is like hockey sticks, like you guys are— hopefully used to seeing. In log space, it's just a nice straight line. This is compute. These are computers getting faster and faster, Moore's Law. There's also data, more and more data available, natural language processing and other, other domains. And this increase in compute and data kind of directly has led to better and more powerful models. This has been a little bit surprising over the last decade or so, but we found that basically faster computers and more data kind of directly translate into just better ML, better AI. We've seen it in a couple areas, so I'll give you a couple quick examples. This is an algorithm from DeepMind called AlphaGo that beat the world champion in Go a couple years ago. A lot of people thought Go is this game that would be so hard to beat, for AI to beat humans, it'll take until 2050 or something. And here we are. This is 2016, I guess this was published. Here's an example from my own work. This is a little robot in our lab. I'm going to turn the sound on. I don't know if that'll work. We had this little 3D-printed robot in our lab. So this is kind of cool if you look carefully. There's 9 servos. So like 1, 2, 3, 4, 5, 6, 7, 8, and there's a 9th one in the middle here, which is like a hip joint, kind of. The whole robot is 3D-printed. So you just like click print, and like a day or two later, out pops this crazy assembly of plastic, and you screw it together with some screws with off-the-shelf servos. And then we want to make this thing walk. We want to make it like go in some direction. So, So the first thing we tried is, like, hand-coding a gait. This is a gait hand-coded by a grad student. And you can see, like, if you imagined how you should make this walk, you might imagine it kind of, like, leans this way, leans this way, back and forth, back and forth. And it is kind of slowly making progress, right? But of course, we do machine learning. We want to make it learn to walk by itself. We tried a shallow neural network, not to go into too many of the details, and eventually learned to kind of move its limbs a little faster. It's walking a bit, but it's not moving much, much faster. We switched to a different type of neural network called— well, it's a simple network that paints a recurrent neural network using an algorithm called HyperNeat. And using this algorithm, we were able to get much more organic motion. So here the limbs are kind of moving together, but it's not really going anywhere. So we thought this was like a step in the right direction. It represents some progress. But it's not like kind of vaulting over the floor like we hoped. Here's— sorry, here's after a bit more work. That's the same robot from before, but now it's experiencing the world. It's learning. It tries something. We actually measure its position via this little LED on top. And every time it moves, we tell it how far it moved, and it learns to move further and faster. At the end, we finally found this gait here where the robot's kind of like wandering like this, pretty fast, pretty coordinated. This gait is 9 times faster than the gate I showed you in the beginning that was coded by a human. So robotics, uh, natural language processing— actually, there's a huge amount of progress being made recently, motivated by a couple bits of work. In particular, the transformer models from Vaswani et al. in 2017 from Google, and then some folks from OpenAI kind of took these to the next level, trained the biggest one at the time to date. And these language models, basically the AI kind of reads Reads like millions and millions of web pages, gets a feel for how humans tend to speak, and then it has a model. It has a generative model for what text is natural. So for example, one thing you can do to this is you can say, I'm the— as a human, I'm going to type in this prompt. In a shocking finding, scientists discovered a herd of unicorns living in a remote valley in the Andes, and so on. I'm going to type this prompt. I'm going to give it to the model. I'm going to say, hey AI, If I started a sentence or a web page like this, how do you think it would end? So I typed this in, or rather the researchers at OpenAI typed this in, and they asked the model to complete it, and the model spits out all of this stuff. The scientists named the population after their distinctive horn, Ovid's unicorn, dot dot dot. Down here is something I like a lot: Dr. Jorge Perez, an evolutionary biologist from the University of La Paz, dot dot dot. So what did the model do? What did this ML do? First, it invented a name for them, right? They named them Ovid's Unicorn. That's pretty cool. Ovid's Unicorn wasn't in this, like, prompt text at all. So, like, these models can be creative. Second, I like this, like, little detail about Jorge Perez. So it makes up a fake doctor, a fake biologist from the University of La Paz. Why is that kind of cool? So University of La Paz here is, It's like, it's a university that plausibly would be in the Andes Mountains. And so somehow this model, just by reading a bunch of web pages about random stuff around the world, learned that, like, in the Andes there might be a biologist named Jorge Perez at University of La Paz. So these types of models are really, really starting to change the world. Thank you.
