About the talk
At the recent AI Summit in San Francisco, Zoubin Ghahramani, chief scientist at Uber and professor of information engineering at the University of Cambridge, presented a session dedicated to probabilistic machine learning - the advanced approach to ML that can suggest the correct result even when the necessary data is not readily available.
“Having been an academic for about 20 years, I’ve seen many of the trends in machine learning and AI come and go, it’s a very fad-ish field of research,” he said. “Of course, the big one that everybody talks about is Deep Learning.
“But we’re reaching the point where we have milked this technology quite a lot, and we need to look out for what’s next.”
Keep up-to-date with all the latest AI news and insight at http://www.aibusiness.com
Welcome to Nai business TV here at the AI Summit San Francisco 2019. I'm very pleased to be joints day vitamins Armani. Who is the chief scientist at Hi. How are you doing? I'm doing fantastic stage from the keynote presentations this morning. You haven't been able to join us today at the ice limit. What were the kind of key takeaways from that thing is that when people think about AI an Uber they often think about autonomous vehicle, but actually a i broadly defined Powers a lot of Step or services that Uber delivers
and so the way I would explain that is although a i is a term that people don't always understand. Sometimes they're influenced by science-fiction when they think about AI very prosaically to think about AI as taking data and using algorithms to make predictions and to optimize decision and that actually underlies everything that we do do Uber. So it's very much a note living in the way we Model cities we predict ETA we optimize matches to reduce interactions for customer service. We provide
safety for our customers. We make the platform as seamless as possible for all our users. We get better location data, all of these things are actually a i driven lessen the core part of the business machine learning how to break through the Innovation. What are the interesting areas where a I am machine learning can help tremendously is in safety. So Uber is a platform that connects Millions of people and when you have a system like that safety is incredibly important and we can use signals and information and machine learning algorithms to provide all
sorts of things to make with Service as safe as possible. And so that's one of the areas where our team has been very heavily involved. For example of one of the things that was launched in our by our CEO and safety launch recently was hands-free pick up experience for drivers. So we really don't want drivers to have to interact with the screen on the mobile phone more than they need to sell conversational speech interfaces is something that powers that
interaction with the many many examples like that including like, you know reaching out to people If they may have been an accident reaching out to them in real-time very quickly and sort of offering to connect them to emergency services including as well like alerting to situations that may be unsafe for the rider or the driver. So these sort of Science Fiction is a robot car that's going to drive you from A to B, but they're very very important to our millions of customers and Driver Partners massive
strike them out. So it's almost full for granted. It's on everybody to be on the happy path. So the happy path is that magical experience where you know, you know, whether you're a rider a driver a restaurant a courier driver on Uber Freight all of these different. Customer bases these people that we interact with we want their experience to be smooth. And so there many things that can take you off the happy path everything from you know, your location data is bad, right and you know this you get picked up in the wrong place and you know that can beat your
safety problems as well. You know, maybe you're unnecessarily going to cross the street when you shouldn't have to cross the street things like that so they may sound trivial but when you multiply these things by the the literally tens of millions of experiences that people have in the real physical world, these are things that we need to get right and we use AI machine learning algorithms to really try to get them as right as possible. And of course it's the work is never finished so we can always do better. They took about the work never been finished. Then I'm here up
a Pioneer with a machine-learning would have been kind of the biggest rides that is has happened in the industry. And what kind of see Yet to Come what's going to be the next to a year or so I've made an interesting transition being an academic for most of my career to working in industry in the last 3 years and you haven't been an academic for about 20 years. I've seen many of the trends in machine learning and AI come and go field research. Of course, the the big one that everybody
talks about his deep learning. These are the machine learning systems that people tend to associate with like the way the brain works and things like that and you hear about that a lot in the popular press it turns out and there's nothing really They're just they're almost like Route 4 systems that extract patterns from massive amounts of data. There has been a tremendous amount of progress that has been made in the last couple of years because of deep learning but it's not just because of deep learning really
the things that are driving that are the availability of massive amounts of data through the internet availability of massive computing power through gpus is a graphic processing units that were originally developed for gaming but have ended up being like a major Hardware driver for the AI industry. And so these Trends Mastiff compute compute massive amounts of data massive number of people coming to the field and deep-learning have caused a lot of Real breakthroughs to happen. So that's a good thing. But we're reaching the point. I feel where we milk this,
you know technology quite a lot and we need to look out for what's next. What's next on the horizon and so will some of the things that are next on the rise in our areas that we've invested in Uber along with many other companies on these include areas where for example you really want to automate things much more. So so machine learning ironically is a very handcrafted thing so is very Talent intensive. So you need to find experts to tweak algorithms and data sets
massage, actually a lot of that could be automated. So this is the area of automated automl Auto Machine learning. That's an area that we care a lot about as well as the classic deep learning methods. Are terrible at representing on certainty and uncertainty is just the fact that when we're trying to predict what's going to happen with you data. We don't really know and we want our machine Learning Systems to know when they don't know. It's a signal when they don't know and you can imagine for many many applications like you-know-who
fooding application template for safety. You really want systems that know when they don't know and so that's an area probabilistic machine. That's an area that carries very old ideas from cystic when I say very old. I don't mean a decade. I mean from the 1700s Mary those ideas with modern deep learning method to get models that are more honest about their lack of knowledge. One of our kind of most recognizable figures kids. Today's living Yorkie nice and
I've had an unusual background. I mean in the sense in a couple of different researcher when I was about thirteen or fourteen, so I've been incredibly consistent on the one hand. I did have some detours along the way at some point. I didn't really know whether I wanted to become an AI researcher or become a computational neuroscientist. So computational Neuroscience is somebody who uses mathematics engineering principles algorithms to understand how the brain works that relates more to biology. And so I went down that path for many years. I did my PhD in computational
Neuroscience actually and then I couldn't Choose population on Earth science, which is the field of science or AI which is more engineering you build thing up until the point that I realized. I wasn't very good as a publication that was much better suited for doing Ai and machine learning and then I made that choice then I thought I was going to be an academic for my whole career and I was wrong there. I also thought that the field of AI and machine learning was you know, intellectually very interesting like particle physics is interesting or you know,
fundamental biology is interesting, but I didn't realize that it was going to become so practical and what I mean by practical is Is an anecdote. I remember flying into San Francisco and going probably in an Uber from the airport to downtown and seeing a billboard. Saying machine learning and I just thought that's that's bizarre because I've been in this field for 20 years. I never thought that the name of that field was going to appear on Highway Billboards dry and that's how how much the field is really impacted the real world as we say in Academia. And so so
I was in academic but as the AI industry started picking up there were a huge number of startups and I got involved along with some of my students and colleagues in the number of different startups in advisory roles. Originally, then I was involved in founding of a startup and and then at some point that start up getting the interest of many of the big tech companies and Uber was as the perfect spot for that in the sense that some of the larger tech companies and already made big AI investment so our team and our
technology would have Just absorb into some larger research Community know they want a I really don't know what to do with it. So they will acquire some AI technology or build an AI division or acquire an AI startup and they don't know how to integrate into their business. You know, there is an organ rejection sometimes like those those efforts fail Uber knew exactly why I needed for business. Is it already made a major investment in autonomous vehicles and I really push the whole Technology field of autonomous
vehicles to another level and when I came on board with my team, we got situated in the Uber headquarters on the main floor. Where the sea Doing all the other Executives were deeply integrated into the company and this is really part of the I would say part of the recipe for success is at their number thing. One of them is be bold and integrate deeply into the business recruit talented people and to do that. You have to recruit some senior town to people people and then more Junior talented people will come join you and then also,
The business focus but be flexible because researchers in the field of AI and machine learning have a tremendous number of options these days and so they they need to be able to engage with the external Community. Even if there an industry they need to be able to publish they need to be able to talk about some of their work. They need to be able to attend conferences open source software whatever their colleagues are doing at the other major tech company. So companies is triway from the basic, you know, tricks end
up not having very successful day. I came so it's a tremendously collaborative industry. So People share ideas. They published papers. They open source is really revolutionize the way we do things. So we need to all engaged in this effort. Thank you so much for today.
Buy this talk
Access to all the recordings of the event
Buy this video
With ConferenceCast.tv, you get access to our library of the world's best conference talks.