Andrew Ng at Amazon re:MARS 2019

Please welcome CEO of landing AI andrew ng. Hey. Hey everyone. It's nice to see so many of you. Thanks for coming to this. I've been saying for four years now that AI is the new electricity. And since then look how much progress AI has made and how much momentum we've sustained. Look at the number of jobs we've created. According to the recent AI index, the number of jobs in deep learning over a two year period increased 35 fold. That's 35x. And look at the number of research papers skyrocketing. And this affects not just the number of papers being published, but also the size of the research community. There's now doing work on AI and publishing all these papers. Or look at the number of mentions on AI of on earnings calls. Whereas the last two slides I think are good evidence. This one, I don't know if you want to take anything from this one, but if you look at what we as a community have done, think of the great AI companies, you probably think of a lot of the software Internet companies like Amazon, Google, Baidu, Facebook, Microsoft, maybe Apple. Um, and what I hope you take away from this presentation, the one thing I hope you take away is that this is the time to bet on AI's rise outside the software industry. And by that I mean companies whose primary business is not selling a software product or running a website or running an app. And Verna Vogel alluded to encouraging signs of this across many companies as well. But we still have a lot of work to do. Let me tell you why. Let me explain further why I think this is important. McKinsey estimated $13 trillion of global GDP value creation by the year 2030 due to AI. But if you look at the industry sectors where this will take place, a lot of it is in sectors like retail, transportation, logistics, manufacturing, agriculture. These are all primarily sectors outside the software industry. We as a community would transform the software industry. Both the large software companies and the small software companies, they get it. But we now need to go and take AI to the rest of the industry, which is actually much bigger than the software industry. Now AI is also more accessible than ever before. You've heard the story of how data and compute has been driving the rise of AI. And this is true, it has. I remember in the early days of when I was leading the Google Brain project, building out the company wide deep learning infrastructure. One of our first project was this slightly infamous Google Cat where we used 16,000 CPU cores in order to learn to recognize cats from YouTube videos. And this made the homepage the Front page of the New York Times website. And this was a big deal back then. And even within Google, it was a little bit of a struggle, even with Jeff Dean's help, to get that many computers today. Any of you can go home, log into AWS, enter your credit card number, and for $2,000, which isn't nothing but $2,000, you can replicate this result. So the rise of data and compute has driven a lot of the rise of deep learning. But the story is now even bigger than this. And the rise of talent, ideas and tools is making AI more accessible than ever before, even to companies outside the software industry. Let's start with talent. Many of you know that several years ago I put my course, my machine learning course online. And whereas I used to teach 400 students a year at Stanford, since then, over two and a half million learners have enrolled in my online machine learning course. But even bigger than that, today universities around the world are teaching more machine learning and AI and deep learning than ever before. And even many corporations have internal educational programs. Many of these people that are taking these courses are highly motivated, using their own spare time, right when doing something else. They're studying this stuff at home in their spare time. Many of them are well skilled, highly trained, highly motivated, and can tackle many problems in many industries. So talent is more available than ever to every industry. Second, the rise of ideas. I remember when about 20 years ago I would go to the machine learning conference. The ICML International Conference of Machine Learning helped roughly every summer. And my community would publish about 150 papers a year. So I would go home and I would literally flip through the paper proceedings and I would read every single abstract of every single paper published by my community. I saw Jeff Dean just tweet last Sunday that the number of machine learning papers on Arxiv now surpasses 100 papers per day. And we've reached a point where no human can keep up a very single paper publishing machine learning. And this is a sign of the sheer volume of ideas that are now out there on the Internet, free for anyone to download and use. Finally, the rise of tools. Software platforms like this is making it easier than ever before for anyone to apply machine learning. A few weeks ago I was attending Maker Faire in California. Maker Fair is a big expose of lots of robotics and maker projects. And I met this kid that had traveled from India to present his robot. He built a robot that was taking pictures of plants, you know, in a, in a field and using machine learning to classify if the plants had a disease. Or not, and if so, what disease? So I asked him, how old are you? And he said I'm 12. And it occurred to me that if five years ago, if a Stanford PhD student had done this project, it would have made a strong chapter in their PhD thesis. And now we have a 12 year old kid from India building on open source tools able to do this type of work. The increased accessibility of AI machine learning is important because with the rise of data and compute, but also talented ideas and tools, this means that AI is the purview, is accessible not just to the elite software companies, but is accessible to every company, many companies all around the world. Now when I speak with executives, especially outside the software industry on AI adoption, the number one challenge I hear is they say, hey Andrew, how do I find the right use cases? Because AI isn't some SaaS product that you can log on to a website and just buy AI for your company. AI will change the core of how many businesses operate. So. So how do you find the right use cases? I want to share with you three ideas I've learned on how to do this effectively. First one, start small. I've seen a lot more teams fail by going too big than going too small. And there's one specific lesson I want to learn from the early days of Google Brain that I want to share with you because I think it'll be useful for many companies, which is that in the early days of building out deep learning at Google, there was still widespread internal skepticism about deep learning. Much as today in a lot of companies there is still some internal skepticism resistance to AI. So my first internal customer at Google was the speech team. And speech recognition is a is a nice project, but it's not the most important or lucrative project, right? It's not web search or advertising. But by working the speech recognition team and making Google's speech recognition system more accurate, that helped other teams in Google start to see what deep learning could do. And it helps us start to gain allies and people that believe in us. My second internal customer was the Google Maps team where we used computer vision to read house numbers to help more accurately geo locate houses on Google Maps and improve the quality of maps data. And it's only after successfully delivering these two that there was enough faith in us, there was enough momentum that I then started the conversation with the Google Ads team. So for many companies too, my advice is start small, find a three to six month project, do it quickly, gain momentum and then build from there. Second, automate tasks, not jobs. I found a useful framework for brainstorming machine learning projects to think of tasks for automation. Now in society, there's been a lot of discussion about AI automating jobs. And this is good. This is important that society has this debate because, because this is this, this will have a big impact, but there's a framework for brainstorming AI use cases. I found it more useful to look at jobs which comprise many different smaller tasks, and then look at the tasks that make up a job and see what's amenable for AI automation. For example, take the healthcare industry. Radiologists do a lot of things. Radiologists read X ray images. They also consult with patients, they do surgical planning. They also mentor younger doctors. So radiologists actually do a lot of things of all of these tasks that the radiologist does in their job. One of them reading X ray images and seems per amenable to AI automation acceleration. And that's why many teams, including mine, are building systems like these to try to enter have AI enter this task of reading X ray images. Third, combine AI and subject matter expertise. There is a certain set of things that AI machine learning, deep learning can do. There's a certain set of things that are valuable for your business and for the use cases. It's important to choose projects at the intersection of both of these two sets. Unfortunately, today only AI people or AI experts people knowledge about AI have a good sense of what's in the set on the left and only subject matter experts, people that know your business have good sense of what's in the set on the right. So I recommend to most companies to form cross functional teams or to form teams where you put AI experts and subject matter experts together and have them sit on the same team and brainstorm projects together. I often recommend to companies to brainstorm at least six projects and for each of them to deeply evaluate whether or not it is something that AI can do and whether or not or how much it is actually valuable for your business. And only after doing that for say half a dozen projects, do you select one or two to invest in. One pro tip. I've also found that often the number one project that a CEO suggests is not the right one to invest in. Let me tell you one more story about why it's important to work with subject matter experts. About a year ago, I was visiting a factory in a very remote city in China and it had not snowed for over 10 years in that city. The last time it snowed over 10 years ago, the plastic parts that this statue was making came out melted. And all funny because it was really cold. So I asked them, what will you do if it ever snows again? And they told me, they said, well, Andrew hasn't snowed for 10 years. We're not sure what we'll do if it ever snows again. But those three guys, they were here 10 years ago and said, if it ever snows again, we're going to do whatever they tell us to do because we think they will know best. So in the software industry, it's easy to forget how much other businesses are run based on tribal knowledge and intuition and experience. And I think it's important for our AI teams to get together with people such as those three guys to work together to find meaningful use cases. Also, if we can use machine learning to make those three guys work a bit more systematic, then this means that if any, if they ever retire, that the factory will be able to continue on in a more systematic way. So to summarize, we talked about starting small automating tasks rather than jobs and combining AI and subject matter experts. And I hope that with these suggestions that you really hope many companies, not just software companies, many companies outside the software industry find great use cases. Lastly, I'd like to leave you with a few upcoming AI ideas and upcoming AI technologies that are not widely used, even, maybe not even widely known today, but I think will have an important impact on products and businesses in the future. So here are my top three. First, AI will become a systematic engineering discipline. And I think to some of you this will sound boring, but this is important. Make an analogy. In the days of yore when we were building bridges, there used to be these wiser men and wiser women that would say, hey, add an extra piece of wood there. And if you did what they said, your bridges were more likely to stay up. And it took us a long time to evolve to the systematic engineering discipline of civil engineering, so that teams all around the world are now much better at building bridges that do stay up. AI is still in the process of evolving from this wisdom of wise old men and wise old women to being a systematic engineering discipline. And this is important because we need everyone all around the world to know how to build this in a systematic way. To make another analogy, even in the software industry, it took us a long time to figure out processes to help software teams work together effectively. It took us a long time to figure out ideas like engineering agile development with two week sprints or code review or version control, right? These are actually really important ideas. And the field of AI is still earlier and we're still figuring out the right processes to help teams work together effectively to build machine learning systems. One concrete example, a lot of my machine learning teams use not two week sprints, but instead one day sprints in which my teams would write software during the day, run an experiment overnight, and then the next morning see the results from last night, and then decide the next morning how to reprioritize and what to do. And again, to some software teams this is weird. You do a one day sprint. But teams like Landing AI and many others are innovating on the processes to figure out how to get AI teams to work together more systematically. Second, we need to transition from big data to small data. And if you can get tons of data, I love it, you know, more data, great. But a lot of the rise of AI has been led by consumer Internet companies which have the luxury of having a lot of customers and does a lot of data. Once you look outside the software industry, a lot of companies just do not have these giant data sets. But we need AI to work on small data so that it can apply to them as well. For example, I've been doing a lot of work on visual inspection or surface defect inspection. Today, when a smartphone rolls off a factory floor, the factory needs to inspect it to see if it is scratched or not. Now, if you have a million pictures of scratched cell phones, then you know it's not that hard to build a convolutional neural network to recognize scratched cell phones versus good cell phones. But fortunately no factory in the world has ever manufactured a million scratched cell phones that then need to be thrown away. This data just does not exist. So the question is, if you have only 100 pictures of scratched cell phones, can you still get your neural network to work? Thanks to improvements in technologies such as few shot learning, transfer learning and data augmentation, we're making a lot of progress in small small data. And this would break open a lot of new applications of AI to a lot of industries outside software where the data sets are often much smaller. Finally, one last technology I'm excited about is meta learning for reinforcement learning or meta reinforcement learning. You know, reinforcement learning's height PR hype has been vastly disproportionate relative to its actual commercial impact today. But robotics is but I find robotics exciting. We heard great talks from Jeff Wilkie, Mark Rabut, Kate Darling over the past couple days of various robots and think of the robots that in the future will be the nimbly assemble intricate electronic devices like cell phones or in agriculture. Think of the Robots as starting to pick delicate strawberries. Or for space exploration, we will need robots with a high degree of autonomy because of the latency of communication between Earth and other planets. So let me share with you this niche idea, this idea that maybe isn't widely known today, but I think will have a big impact in the future. So to explain this idea, I used to fly autonomous helicopters at Stanford for a long time. And the way we would use AI to control it is we will always build a simulator, basically a video game of the helicopter, and have the AI learn to fly in simulation and it will copy the software over to the real helicopter. Now, the problem with this is that no one can build a perfect simulator or a perfect model of the physical helicopter. So with meta learning, which didn't exist back then, but if it had, this is what we could have done, which is, Rather than build one simulator for a helicopter, let's build five or 100 different simulators for helicopter, where the simulators all vary a bit. Maybe in one this heavier, the helicopter is heavier, or the different one helicopter is lighter, another one stronger thrust, different dimensions, but had this portfolio of similar status that all vary a little bit. What the AI's job is to do is to be able to wake up in any of these simulators, and whatever simulator suddenly wakes up in this job is to quickly adapt to fly that one helicopter simulator well. So researchers have found people like Peter Bio, Chelsea Finn, others at UC Berkeley, Google, OpenAI, DeepMind have found that with this type of process, if you take the RL algorithm that's learned to do this and apply it to a physical robot, then this learning algorithm will be much better at adapting to whatever other dynamics of the actual physical robot you have. Um, and I think this will have a big impact on a lot of robotics applications, including space exploration, where no one really knows how to build an accurate simulator, you know, some of the extra effects. So we've talked about how AI is more accessible than ever. We've talked about how to find the right use cases, as well as some upcoming technologies I think will have a big impact on products and businesses. And with these, I hope we can take what we've done transforming the software industry, but also transform the rest of those industries that need our help and are actually much bigger than the software industry. I want to close with one final thought, which is that in eras of technological disruption, leadership matters. For those of us in the software industry, even as we led the world through multiple waves of technological disruption, we've not always provided the best leadership with the rise of the Internet, we created tremendous wealth, but we also contributed to wealth inequality. That's why I think now, with the rise of AI, it is incumbent on all of us to invest at least equally, if not more, on antidotes such as education, or making sure we have informed government leaders and thoughtful regulation. Because let's make sure that this time, with the rise of AI, even, even as we're building this exciting AI powered future, that we take everyone along with us. Thank you very much.