Mike Schroepfer chats “AI 101 with Joelle Pineau” co-lead of the Meta (Facebook) AI Research Lab •

In this AI video ...

Hey everyone, welcome to this week’s Shreptech. I am unbelievably excited about today’s guest. We have Joel here who leads our Montreal Lab in a large part of our Facebook AI Research Lab. She’s also a professor at McGill University, and she’s going to talk to us about AI. Joel, welcome to Shreptech. Thank you. Really happy to be here. Okay. So I have this theory that anyone can learn AI. If they put their mind to it. So let’s just start. If I’m a newbie and I don’t know anything about AI, tell me about it. What is it? It’s interesting. It’s been evolving over the years. The concept of AI itself started around the 1950s, though it had been in the imagination before that, with automotons and robots and so on. But from the early days, it took AI and split it into a bunch of sub-disciplines. So can we understand language? Can we understand images? Can we form memories? Can we perform association between objects? It proceeded with these different fields for several decades until one of these pieces sort of emerged as the dominant one. The most important part of the puzzle, which is machine learning. A lot of time nowadays when people talk about AI, they actually report to machine learning, which is the ability to digest large amounts of data and use this data to essentially make predictions. So one way to think of AI and the most common way we use it today is as a little prediction machine. You can imagine that if you have this prediction machine, so as close as we get to a magic eight ball for computer science, you can put that to lots of really important uses. Give us an example. What are some real world examples of this prediction machine? This is something that we’ve used for example to analyze images. So you take an image of something and this little prediction machine can tell you what are the objects in that machine. It can tell you some of the interaction between the object. It may even be able to take some questions about the object. So is Joelle on the left or on the right of by looking at the overall view? Okay, so if I’m trying to describe the difference between AI and the subset that is ML or machine learning, how what’s a good way to explain that difference? Think AI is really the broader picture of just intelligence with an analogy to human intelligence. It’s really quite grounded in notions of human intelligence. And so it’s really the ability to reason, to plan, to form memories, and so on. But machine learning is often the part that sort of fuels all of these other functionalities. Just like in humans, the ability to learn really opens up our ability in terms of all the other cognitive functions. Another common question or mistake I see people make is they see a press of result in one area. So years ago, AI systems beat grandmasters ago. And you say, well, wow, I can’t beat grandmasters ago. So that AI could beat a grandmaster. It must be able to do lots of things. But can it? Well, many AI’s together. If you have your toolbox with, but you’ll need one algorithm or one learner who’s learned to play Go, and another one who’s learned to write poetry and yet another one who’s learned how to compose music, having a single model that’s trained to do all of these things is still quite far away. And in particular, most of the time, the AI systems we have are best at digesting really homogeneous data. Images, of course, two images might look different, but at the end of the day, it’s a bunch of pixels. The language of the image is a pretty restricted language. So each machine learning model is really best when you feed it only one type of information. Mixing between types of information, and we try to do it, for example, when we work with medical data, where there’s all sorts of tests and labs and demographics and so on, it’s a lot harder. And that’s where we really still have a lot of progress to make, including with multimodal data. Well, so here’s another question for you, which is, you know, the field is moving so fast. And even if we shrink, you know, so we started this with, okay, there isn’t like an Uber brain that’s sucking up all this information and capabilities. We’re sort of building these more specialized tools. It’s a hammer and a screwdriver and a saw, and they’re really good at their task and language and video and whatever, but they don’t sort of accumulate to some master craftsman or master brain. The other thing, though, is that people miss the rate of progress, as you just talked about in the field of ML and machine learning. And so someone who maybe tried something a decade ago, said, well, I tried this whole natural language processing thing, and it doesn’t really work. And what, you know, what it kind of does now, what’s changed, what’s new, what would an out-of-date practitioner miss because they miss the state of the art of the field right now? Hey, I will confess that the researcher, there’s things we tried six months ago that didn’t work, that all of us on our working. And so it’s really changing. There’s a few things that are changing, usually, you know, three ingredients. One is we didn’t have the data, the scale of the data, the variety of the data to show that it worked. The other one is we didn’t have the compute. So we weren’t training a model. You can think of a machine learning model as like a very complicated mathematical function. We weren’t trying a function that was complex enough because we didn’t have the compute infrastructure to do it. And the third one is we didn’t have the right set of mathematical formulas in our function. So the right model to do it. And I would say, you know, progress on all these three things has been growing really quickly over the last decade. And so absolutely, there’s things we couldn’t do very shortly. And it’s usually a mix. You know, more data needs more compute. And once you’re in a different regime of data, there are different models that are necessary. You know, right now for the last several years, everyone’s super excited about neural networks as being the tool to digest very large amounts of data, very large amounts of computation. Many people come to me wanting to use these neural net, but there’s lots of applications where that’s not the right tool. When you’re in the small data regime with different characteristics, there’s lots of other tools in your toolbox that you should be using. So you have to be really mindful of that combination. Well, what else do you wish more people understood about AI and ML? There’s a lot of people who think of AI and ML, in particular, as a black box. I keep on hearing this. It’s kind of a black box you put it in. You have no idea what’s going in there. And I wish more people understood that it’s a very complicated box. There’s a lot of things going on in that box. But in many ways, it’s much more, you can audit what’s in the box. You can rerun what’s in the box. You can retrain it. It’s a lot more understandable. Of course, once you know the language of machine learning, but it’s a lot more understandable than what’s going on in human intelligence. People think, oh, humans are so rational and these machine learning things are completely incomprehensible. Humans are very good at rationalizing their thinking process. The machines much less. And so one of the things we’re doing is really helping the machines explain their decisions so that they can communicate with humans. And in many really safety critical applications, if you think of use of AI for medicine, for the legal system, and so on, having that visibility, that understanding is really important. So not a black box, but a very complicated one.

AI video(s) you might be interested in …