Following is the full transcript of the entire Google I/O 2019 developer keynote event. Google’s CEO Sundar Pichai and the team announced latest products and services that the company provides. This event occurred on May 7, 2019 at Shoreline Amphitheatre, Mountain View, California, United States.
Speakers at the event:
Sundar Pichai – CEO, Google
Aparna Chennapragada – VP of Product for AR and VR, Google
Scott Huffman – Vice President, Google Assistant
Stephanie Cuthbertson – Senior Director for Android
Rick Osterloh – SVP of Hardware
Sabrina Ellis – VP of Product Management
Jeff Dean – Lead of Google AI
Lily Peng – Product Manager, Google AI Healthcare Team
Sundar Pichai – CEO, Google
Good morning. Good morning. Wonderful to be back here at Shoreline with all of you.
It’s been a really busy few months for us at Google. We just wrapped up Cloud Next in San Francisco with over 30,000 attendees, as well as YouTube Brandcast last week in New York.
Of course, today’s about you all, our developer community. And thank you all for joining us in person, and to the millions around the world watching on livestream.
I would love to say welcome in all our languages our viewers speak, but we are going to keep the keynote under two hours, especially since Barcelona kicks off against Liverpool at noon for you. That should be an amazing game.
Every year at I/O, we learn and try to make things a little bit better. That’s why we have lots of sunscreen — hope the sun comes out — plenty of water and shade. But this year, we want to make it easier for you to get around. So we are using AR to help.
To get started, open your I/O app and choose Explore I/O. And then you can just point your phone where you want to go. We really hope this helps you get around and answers the number one question people have: where the sessions are. Actually, it’s not that. They want to know where the food is. And we have plenty of it around.
We also have a couple of Easter eggs, and we hope you enjoy them as well. This is a pretty compelling use case. And we actually want to generalize this approach so that you can explore and navigate the whole world that way. There’s a lot of hard work ahead. And it’s a hard computer science problem. But it’s the type of challenge we love.
Tackling these kinds of problems is what has kept us going for the past 21 years. And it all begins with our mission to organize the world’s information and make it universally accessible and useful.
And today, our mission feels as relevant as ever. But the way we approach it is constantly evolving. We are moving from a company that helps your find answers to a company that helps you get things done.
This morning, we’ll introduce you to many products built on a foundation of user trust and privacy. And I’ll talk more about that later.
We want our products to work harder for you, in the context of your job, your home, and your life. And they all share a single goal: to be helpful, so we can be there for you in moments big and small over the course of your day. For example, helping you write your emails faster with automatic solutions from Smart Reply, and giving you the chance to take them back if you didn’t get it right the first time, helping you find the fastest route home at the end of a long day, and when you get there, removing distractions so that you can spend time with the people most important to you.
And when you capture those perfect moments, backing them up automatically so you never lose them.
Simply put, our goal is to build a more helpful Google for everyone. And when we say “helpful,” we mean giving you the tools to increase your knowledge, success, health, and happiness. We feel so privileged to be developing products for billions of users. And with that scale comes a deep sense of responsibility to create things that improve people’s lives.
By focusing on these fundamental attributes, we can empower individuals and benefit society as a whole. Of course, building a more helpful Google for us always starts with search and the billions of questions users trust Google with everyday. But there is so much more we can do to help our users.
Last year, we launched a new feature in Google News called Full Coverage. And we have gotten great feedback on it from our users. We’ll be bringing Full Coverage directly to search to better organize results for news-related topics. Let’s take an example.
If you search for “black hole,” we’ll surface the relevant top news. It was in the news recently. We use machine learning to identify different types of stories and give you a complete picture of how a story is being reported from a wide variety of sources. You can click into Full Coverage. It serves as a breadth of content, but allows you to drill down into what interests you.
You can check out different aspects of the story, like how the black hole got its name. You can even now see a timeline of events. And we’ll be bringing this to search later this year.
Podcasts are another important source of information. And we’ll be bringing them directly to search as well. By indexing podcasts, we can surface relevant episodes based on their content, not just the title. And you can tap to listen right in search results, or you can save an episode for listening later on your commute or your Google Home.
These are all examples of how we are making search even more helpful for our users, surfacing the right information in the right context. And sometimes, what’s most helpful in understanding the world is being able to see it visually.
To show you how we are bringing you visual information directly in search, here’s Aparna.
Aparna Chennapragada – VP of Product for AR and VR, Google
Whether you’re learning about the solar system or trying to choose a color scheme for your home, seeing is often understanding.
With computer vision and augmented reality, the camera in our hands is turning into a powerful visual tool to help you understand the world around you.
So today, we are excited to bring the camera to Google search, adding a new dimension to your search results — well, actually three dimensions, to be precise. So let’s take a look.
Say you’re a student studying human anatomy. Now, when you search for something like muscle flexion, you can view a 3D model built by Visible Body right from the search results. Pretty cool.
Not only that, you can also place it in your own space. Look, it’s one thing to read about flexion or extension, but seeing it in action right in front of you while you’re studying the concept, very handy.
OK, let’s take another example. Say, instead of studying, you’re shopping for a new pair of shoes. That happens. With New Balance, you can look at shoes up close from different angles, again, directly from search. That way, you get a much better sense for things like, what does the grip look like on the sole, or how they match with the rest of your clothes.
OK, this last example is a really fun one. So you may have all seen a great white shark in the movies. “Jaws,” anyone? But what does it actually look like up close? Let’s find out, shall we?
I have Archana here with me to help with the demo. So let’s go ahead and search for “great white shark” on Google. As you scroll through, you get information on the knowledge panel facts, but also see the shark in 3D directly from the knowledge panel.
Why don’t we go one step further? Why don’t we invite the shark to the stage? Whoa!
There it is. It’s one thing to read a fact like “a great white can be anywhere between 17 feet to 21 feet long,” but to see it in front of you at scale, filling up the Shoreline stage like a rock star, that is truly understanding its scale.
OK, let’s take a closer look. It’s an AR shark. It won’t bite. Ooh. Look at those layers of teeth. You know, I don’t know about you all, but I’d much rather see these teeth up close in AR than in real life.
Thank you, Archana.
Really excited about bringing the camera and AR capabilities to Google search.
Now, sometimes, though, the things that you’re interested in, they’re difficult to describe in a search box. So that’s why we created Google Lens, to help you search and do more with what you see by simply pointing your camera. The built lens has a capability across products. So you can access it directly from the Google Assistant. But we’ve also built it into Google Photos and the Camera app on many Android devices.
People have already used Lens more than a billion times so far. And they’ve used it to ask questions about what they see, like what kind of flower that is, or where to get a lamp like that, or just who the artist is.
One way we’ve been thinking about it is, with Lens, we’re indexing the physical world, billions of places and products and so on, much like search indexes the billions of pages on the web.
OK, today, let me show you some new ways that we’re making Lens more helpful to you. Say you’re at a restaurant trying to figure out what to order. Instead of going from the menu to different apps on the phone and back to the menu and so on, you can simply point your camera. Lens automatically highlights the popular dishes at this restaurant right on the menu.
And of course, if you want to know more, you can tap on any dish on the menu, and you can see what it looks like, again, at the restaurant — and, of course, check out what other people are saying about it on Google Maps.
By the way, when you’re done eating, Lens can help pay for your meal. Not so fast. It’s not picking up your tab. But it can calculate the tip and even split the total — again, just by pointing your camera at the receipt. And voila.
So you saw how we connected the menu with information from Google Maps. But we’re starting to think of other ways that we can connect helpful digital information with the things in the physical world. So I’m going to give you just one example.
So you’re flipping through a “Bon Appetit” magazine and you see a recipe you like. Soon, you can point your camera at the recipe and see the page come alive, showing you how to make the dish. We’re starting to work with more partners, like museums, magazine publishers, and retailers, to bring unique visual experiences like this.
There’s one final area where we think that the camera can be particularly helpful to people. Around the world, there are more than 800 million adults who are struggling to read the words that they come across in their daily lives — bus schedules, bank forms, et cetera. And many of them are coming online for the first time with a smartphone.
So to help with that, we’ve integrated a new camera capability into Google Go. This is our search app for entry level devices. Take this sign in English next to an ATM. Now, for someone who does not understand the language and cannot read the words, this is important information that they’re not getting access to. And we think that the camera can help here. So let me show you how.
So directly from the Google search bar, you can use Lens, open it, point it at the sign to hear the text read out aloud to you.
[Google Assistant: Information for card holders — all customers using old proprietary magnetic stripe cards should be advised.]
What is nice here is that it is highlighting the words as they’re spoken. That way, even if you can’t read the language well, you can follow along, and you understand the full context of what you see.
You can also translate it into your own language, like this.
Notice that the translated text is overlaid right on top of the original sign. It almost feels like the sign was written in your own language to start with. And again, you can hit Listen and hear the words read out loud, this time in your own language.
[Google Assistant: [Speaking Spanish]
What you’re seeing here is text-to-speech, computer vision, the power of translate, and 20 years of language understanding from search all coming together.
Now, our teams in India have been working with some early testers and getting a lot of feedback to make the product better. And I want to now show you how one of them is using it in her daily life. Take a look.
Thank you, Urmila, for testing it and giving us a lot of feedback for the team to make the product better.
The power to read is the power to buy a train ticket, to shop in a store, to follow the news. It’s the power to get things done. So we want to make this feature accessible to as many people as possible. So it already works in more than a dozen languages. And the teams worked incredibly hard to compress all of this tech to just over 100 kilobytes.
That way, it can work on phones that cost as little as $35. So we’re super excited about this and all the other features across Search and Lens to help you throughout the day. You’ll start to see these updates roll out later this month.
Sundar Pichai – CEO, Google
Thanks Aparna. Helpfulness is also about saving time and making your day a little bit easier. That’s why, last year, at I/O, we gave you a first look at our Duplex technology.
Duplex enables Google Assistant to make restaurant reservations on your behalf by actually placing a call. It’s now available in 44 states across the US. And we’ve gotten great feedback not only from our users, but from businesses as well.
For us, Duplex is the approach by which we train AI on simple but familiar tasks to accomplish them and save you time. Duplex was launched with restaurant reservations on the phone. But now, we are moving beyond voice and extending Duplex to tasks on the web.
We again want to focus on narrow use cases to start. So we are looking at rental car bookings as well as movie ticketing. Today, when you make a new reservation online, you have to navigate a number of pages and steps, filling out information and making selections along the way. I’m sure you’re all familiar with this experience.
It’s time consuming. And if users leave during the workflow, businesses lose out as well. We want to make this experience better for both users and businesses. So let me show you how that system can do it better.
Say you get a calendar reminder about an upcoming trip. And you want to book a rental car. You can just ask Google, book a National car rental for my next trip. The Assistant opens the National website and automatically starts filling out your information on your behalf, including the dates of the trip.
You can confirm the details with just a tap. And then the Assistant continues to navigate the site. It even selects which car you like. It’s acting on your behalf and helping you save time, but you’re always in control of the flow.
Let’s go ahead and add a car seat. And once all the details are in, you can check everything one last time and just tap to finalize the reservation. You’ll immediately get a booking confirmation.
It’s amazing to see the Assistant complete a task online on your behalf in a personalized way. It understands the dates of your trip and your car preferences based on trip confirmations in Gmail.
And I also want to point out that this was not a custom integration. This required no action on part of the business to implement. What you just saw is an early preview of what we are calling Duplex on the Web. We’re going to be thoughtful and get feedback from both users and businesses to improve the experience. And we’ll have more details to share later this year.
The Google Assistant helps people around the world with all kinds of tasks, whether they are at home or on the go. But we want to build an even more helpful assistant.
In order to process speech today, we rely on complex algorithms that include multiple machine learning models. One model maps incoming sound bites into phonetic units. Another one takes and assembles these phonetic units into words. And then a third model predicts the likelihood of these words in a sequence. They are so complex that they require 100 gigabytes of storage and a network connection.
Bringing these models to your phone — think of it as putting the power of a Google data center in your pocket — is an incredibly challenging computer science problem. I’m excited to share we have reached a significant milestone.
Further advances in deep learning have allowed us to combine and shrink the 100-gigabyte models down to half a gigabyte, small enough to bring it onto mobile devices. This eliminates network latency and makes the Assistant so much faster — so fast that tapping to use your phone would seem slow.
I think this is going to transform the future of the Assistant. And I’m thrilled to bring Scott to tell you more about our next generation Assistant.
Scott Huffman – Vice President, Google Assistant
Well, what if we could bring the AI that powers the Assistant right onto your phone? What if the Assistant was so fast at processing your voice that tapping to operate your phone would almost seem slow? It opens up many new use cases. And we want to show you how fast it is.
Now, internally, we’ve been calling this the next generation Assistant. Running on device, it can process and understand requests in real time, and deliver the answers up to 10 times faster.
Now, Maggie’s here. And she’s going to help us test it out, starting with some back-to-back commands to demonstrate its speed. Now this demo is hot off the press. So please send your positive energy over in Maggie’s direction.
Maggie: Hey, Google, open Calendar. Open Calculator. Open Photos. Set a timer for 10 minutes. What’s the weather today? What about tomorrow? Show me John Legend on Twitter. Get a Lyft ride to my hotel. Turn the flashlight on. Turn it off. Take a selfie.
Scott Huffman – Vice President, Google Assistant
All right. Now as you could see — Yeah. That was awesome.
Maggie was able to open and navigate apps instantly. Now you might have also noticed that, with continued conversation, she was able to make several requests in a row without having to say “hey Google” each time.
Now, beyond an effortless way to operate your phone, you can start to imagine how the Assistant fused into the device could orchestrate tasks across apps.
Let’s look at another demo where Maggie’s chatting with a friend. He’s going to ask her about a recent trip. Notice how easy it is for her to respond with her voice and even share a photo.
Maggie: Reply. Had a great time with my family, and it was so beautiful. Show me my photos from Yellowstone. The ones with animals. Send it to Justin.
Scott Huffman – Vice President, Google Assistant
All right. Yeah.
Now another example is when a friend asks you a question and you need to look up the information to respond. Justin wanted to know when Maggie’s flight arrives.
Maggie: When’s my flight? When’s my flight? Reply. I should get in around 1:00 PM.
Scott Huffman – Vice President, Google Assistant
So notice how it helped Maggie multitask more easily across different apps, saving her a lot of back-and-forth.
Now you can even imagine this next generation assistant handling more complex speech scenarios, like composing and sending an email.
Maggie: Hey Google, send an email to Jessica.
Hi, Jessica. I just got back from Yellowstone and completely fell in love with it. Set subject to “Yellowstone Adventures.” Let me know if next weekend works for dinner so I can tell you all about it. Send it.
Scott Huffman – Vice President, Google Assistant
Whoa. All right.
Now, as you could see, this required the Assistant to understand when Maggie was dictating part of the message versus when she was asking it to complete an action.
Maggie: Thanks, Scott.
Scott Huffman – Vice President, Google Assistant
By moving these powerful AI models right onto your phone, we’re envisioning a paradigm shift. This next generation assistant will let you instantly operate your phone with your voice, multitask across apps, and complete complex actions, all with nearly zero latency. And actions like turning on the flashlight, opening Gmail, or checking your calendar will even work offline.
Now, it’s a very hard problem we’ve been solving. And I’m really excited to share, the realization of this vision is not far off. In fact, this next generation assistant is coming to the new Pixel phones later this year.
Now our mission is to make the Assistant the best way to get things done. You just saw how we’re making it much faster. But it also has to be personal enough to really help you.
Now, personalized help is especially important in areas where people’s preferences completely differ, like choosing what to listen to, what to do on the weekend, or even what to eat. So let’s look at a recipe example.
Hey, Google, what should I cook for dinner?
Assistant: Here are some recipe picks for you.
Scott Huffman: Now as you can see, the Assistant picked recipes tailored to me. For example, it suggested a bourbon chicken recipe, because it’s helped me with barbecue recipes in the past. Now what I really love is that different people get completely different results. We call this feature Picks for You. And it will be launching on smart displays later this summer, starting with recipes, podcasts, and events.
Now, beyond your preferences, becoming more personal means the Assistant will better understand the people, places, and events that are important to you.
Now, one important person in my life is my mom, who I’m going to visit right after I/O. So let’s say I asked my Assistant, how’s the traffic to Mom’s house? Now, we all understand what I mean by Mom’s house, right?
Well, if I’m in Toledo, Mom’s House might have meant this place, a nonprofit childcare center. In other cities, Mom’s House can be a restaurant or a grocery store. In fact, there’s lots of things in the world called Mom’s House.
Now, in linguistics, the process of figuring out which thing a phrase refers to is called reference resolution. And it’s fundamental to understanding human language.
At Google, we approached this problem using our knowledge graph of things in the world and their relationships. It’s what allows us to understand something like the Starbucks near the Golden Gate Bridge. Today, we’re expanding the Assistant’s ability to understand you better by applying those same techniques to the things in your world. We call it Personal References.
So if I shared my mom’s contact info with the Assistant, I can ask, hey, Google, what’s the weather like at Mom’s house this weekend?
Assistant: Friday and Saturday in Carmichael will be partly cloudy.
Scott Huffman: How long will it take to get there?
Assistant: With light traffic, it will take you two hours and 14 minutes to get to 123 Main Street by car.
Scott Huffman: Remind me to order flowers a week before Mom’s birthday.
Assistant: All right. I’ll remind you on July 3.
Scott Huffman: And it goes beyond Mom. If you’ve shared important people, places, and events with the Assistant, you’ll be able to ask for things more naturally. Like, show me photos of my son, or directions to the restaurant reservation, or reminding me to pick up chocolates on my anniversary.
And rest assured, you’re always in control. You can edit or delete this information at any time in the updated You tab in Assistant settings.
Now, one place where the Assistant can be especially helpful is in the car, offering a safer hands-free way to get everything you need while you’re on the road.
Now, we’ve been focused on the main things that we all want when we’re driving — to get where we’re going safely, to catch up with friends, and listen to something interesting along the way. Last year, we brought the Assistant to Android Auto. And earlier this year, we added it to Navigation in Google Maps. I’m happy to share, the Assistant is also coming to Waze in the next few weeks.
Now, I’d like to show you the future of how we’re improving your mobile driving experience even more. Introducing the Assistant’s new driving mode. Just put your phone in the car and say, hey, Google, let’s drive.
Driving mode has a thoughtfully designed dashboard that brings your most relevant activities front and center while you’re driving, and includes suggestions personalized for you. For example, if you have a dinner reservation on your calendar, you’ll see a convenient shortcut to navigate to the restaurant.
Or if you started a podcast at home in the morning, once you get in your car, it will display a shortcut to resume the episode right where you left off. Now it also highlights top contacts, making it easy to call them or message them, and recommendations for other things to listen to.
Now, once you’re navigating, phone calls and music appear in a low profile way, so you can get things done without leaving your navigation screen. Hey, Google, play some jazz.
Assistant: Sure. Check out this jazz music station on YouTube Music.
Scott Huffman: Now, everything is voice-enabled. So if a call comes in, the Assistant will tell you who’s calling and ask if you want to answer without having to take your eyes off the road.
Assistant: Call from Mom. Do you want to pick it up?
Scott Huffman: No thanks. But thanks for your help with the demo, Mom.
All right, so best of all, with the Assistant already on your phone, there’s no need to download an app. Just start driving. Driving mode will be available this summer on any Android phone with the Assistant.
Now, today, the Google Assistant is available on over 1 billion devices in over 30 languages across 80 countries. And with Duplex on the Web, the next generation Assistant, personalized help, and assistance in the car, we’re continuing to build on our mission to be the fastest, most personal way to help you get things done.
Now, before I go, I want to share a little something that a lot of you have been asking for. Check this out.
Now you can stop your timers and alarms just by saying stop. No “hey, Google” needed. And it’s rolling out on smart displays and Google Homes in English-speaking locales starting today.
Thanks very much.
Sundar Pichai – CEO, Google
Thanks, Scott. It’s great to see the momentum of Google Assistant and how it’s able to help users get things done.
So far, we’ve talked about building a more helpful Google. It’s equally important to us that we do this for everyone. “For everyone” is a core philosophy for us at Google. That’s why, from the earliest days, search was the same, whether you were a professor at Stanford or a student in rural Indonesia.
It’s why we build affordable laptops for classrooms everywhere. And it’s why we care about the experience on low cost phones in countries where users are just starting to come online with the same passion as we do with premium phones.
And it goes beyond our products and services. It’s why we offer free training and tools through Grow with Google, helping people grow their skills, find jobs, and build their businesses. And it’s how we develop our technology, ensuring the responsible development of AI, privacy and security that works for everyone, and products that are accessible at their core.
Let’s start with building AI for everyone. Bias has been a concern in science long before machine learning came along. But the stakes are clearly higher with AI. It’s not enough to know if a model works. We need to know how it works. We want to ensure that our AI models don’t reinforce bias that exists in the real world. It’s a hard problem, which is why we are doing fundamental computer science research to improve the transparency of machine learning models and reduce bias. Let me show you what I mean.
When computer scientists deploy machine learning models, it can sometimes be difficult to understand why they make a certain prediction. That’s because most machine learning models appear to operate on low level features — edges and lines in a picture, color of a single pixel. That’s very different than the higher level concepts more familiar to humans, like stripes on a zebra.
To tackle this problem, Google AI researchers are working on a new methodology called TCAV, or testing with concept activation vectors. Let me give you an example. If it’s a machine learning model, trained to detect zebras, you would want to know which variables were being used to decide if the image contained a zebra or not. TCAV can help you understand if the concept of stripes was important to the model’s prediction.
In this particular case, it makes sense. Stripes are an important predictor for the model.
Now suppose a classifier was trained on pictures of doctors. If the trained data was mostly males wearing coats and stethoscopes, then the model could inaccurately assume that being male was an important prediction factor. There are other important examples as well.
Now imagine an AI system that could help with detecting skin cancer. To be effective, it would need to recognize a wide variety of skin tones representative of the entire population. There’s a lot more to do, but we are committed to building AI in a way that’s fair and works for everyone, including identifying and addressing bias in our own ML models and sharing tools and open data sets to help you as well.
Another way we build for everyone is by ensuring that our products are safe and private, and that people have clear, meaningful choices around their data. We strongly believe that privacy and security are for everyone, not just a few. This is why powerful privacy features and controls have always been built into Google services.
We launched incognito mode in Chrome over a decade ago. We pioneered Google Takeout, which gives you easy controls to export your data, from email, contacts, photos — all of our products — any time you choose to. But we know our work on privacy and security is never done. And we want to do more to stay ahead of constantly evolving user expectations.
We have been working on a significant set of enhancements. And I want to talk you through a few.
Today, you can already find all your privacy and security settings in one place in your Google account. To make sure your Google account is always at your fingertips, we are making it easily accessible from your profile photo. If you’re in search, you can tap on your photo, and you can quickly access the most relevant privacy controls for search, in the case of your data in search.
Here, you can view and manage your recent activity. And you can easily change your privacy settings. Last week, we announced auto-delete controls, which you’ll also be able to access right from the app.
Data helps make search work better for you. And with auto-delete, you can choose how long you want it to be saved — for example, three or 18 months, after which any old data will be automatically and continuously deleted from your account. This is launching today for web and app activity. We’ll be rolling it out to location history in the coming weeks. And we’ll continue to bring features such as this to more controls over time.
In addition, one-tap access to your Google account will be coming to our major products, including Chrome, Search, Assistant, YouTube, Google News, and Maps.
And speaking of Maps, if you tap on your profile photo, in addition to finding easy access to your privacy controls, you’ll find a new feature, incognito mode.
Incognito mode has been a popular feature in Chrome since it launched. And we are bringing this to Maps. While in Incognito in Maps, your activity, like the places you search and navigate to, won’t be linked to your account. We want to make it easy to enter in and out of Incognito. And Maps will soon join Chrome and YouTube with support for Incognito. And we’ll be bringing it to Search as well this year.
Another way we ensure your privacy is by working hard to keep your data secure, from Safe Browsing, which now protects over 4 billion devices everyday, to using TensorFlow to significantly reduce phishing attacks in Gmail, we also encourage users to use two-step verification, because an additional layer of protection is always helpful.
Today, we are making two-step verification even more convenient for everyone by bringing the protection of security keys directly into your Android phone. So now, you can confirm a sign-in with just a tap. And today, it will be available to over 1 billion compatible devices.
We always want to do more for users, but do it with less data over time. So we are applying the same cutting edge AI research that makes our products better and applying it to enhance user privacy.
Federated Learning — this is a new approach to machine learning developed by Google — is one example. It allows Google’s AI products to work better for you and work better for everyone without collecting raw data from your devices. Instead of sending data to the cloud, we flipped the model. We ship machine learning models directly to your device.
Each phone computes an update to the global model. And only those updates, not the data, is securely uploaded and aggregated in large batches to improve the global model. And then the updated global model is sent back to everyone’s device. Let me explain it with a concrete example.
Take Gboard, Google’s keyboard. Using on-device learning alone, when new words become popular, Gboard would not be able to suggest them until you’ve typed them many times. Federated Learning, however, allows Gboard to learn new words like BTS or YOLO after thousands of people start using them without Google ever seeing anything you type.
Actually, with BTS, it’s probably millions of people. This is not just research. In fact, Gboard is already using Federated Learning to improve next word prediction, as well as emoji prediction, across tens of millions of devices. It’s still very early, but we are excited about the progress and the potential of Federated Learning across many more of our products.
Privacy and security are the foundation for all the work we do. And we’ll continue to push the boundaries of technology to make it even better for our users. Building for everyone also means ensuring that everyone can access our products.
The World Health Organization estimates that 15% of the world’s population — over 1 billion people — has a disability. We believe technology can help us be more inclusive. And AI is providing us with new tools to dramatically improve experience for people with disabilities.
For example, there are almost 500 million people in the world who are deaf or hard of hearing. Think of how many conversations are challenging, from in-person discussions and phone calls, to even experiencing videos online.
A few months ago, we launched Live Transcribe, powered by Google’s Cloud Speech API, to caption conversations in real time. You can leave your phone open with the app. And when someone speaks to you, it transcribes their speech into text. Those who cannot — or prefer not to — speak can also respond by typing.
I was really inspired by how the product came about. Two of our Google researchers, Dimitri and Chet, saw an opportunity to help people and collaborated to develop the app. Together, with a small team of engineers and people who volunteered their 20% time, they built Live Transcribe. And it is now available in over 70 languages and dialects on Android devices.
Today, we are going further in extending this technology. We are announcing a new feature called Live Caption. Live Caption makes all content, no matter its origin, more accessible to everyone. The incredible thing is that it works completely on device. So there’s no delay.
With one click, you can turn on captions for a web video, podcast, or even on a moment you capture at home.
[Video clip: You like the blueberries? Blueberries. Delicious? Here comes more. Yum! Show Daddy. Ah.]
It’s only possible due to our recent breakthroughs in speech recognition technology. We recently tested Live Caption with some users. Let’s take a look.
You can imagine all the use cases for the broader community too — for example, the ability to watch any video if you’re in a meeting or on the subway without disturbing the people around you. The Android team is going to talk a little bit later today about what made Live Caption possible.
We’re also exploring how this technology can caption phone calls. But we want to go one step further and actually allow more people to respond and accomplish tasks over their phones.
As you’ll see in this example, Nicole, who’s deaf and prefers not to speak, can receive a call from her hairstylist. With Smart Compose and Smart Reply, she can answer the call and interact. Let’s take a look.
Assistant: Hi. This is Nicole’s Assistive Chat. She’ll see what you say, and her responses will be read back to you, starting now.
Jamie: Hi, Nicole. It’s Jamie. How are you?
Assistant: Hey, Jamie. I’m good. And you?
Jamie: Great. Are we still on for your 1:00 PM haircut tomorrow?
Assistant: Sorry, can you do 3:00 PM?
Jamie: Yes, I can do 3:00 PM. We have a lot to catch up on. I want to hear all about your trip.
Assistant: Perfect. Thumbs up.
Jamie: Great. See you tomorrow. Bye.
Thumbs up indeed.
We call this new technology Live Relay. While there’s still more work to do, we are excited to see how it can help people like Nicole get things done more easily. Just like with Live Caption, this runs completely on device, and these conversations remain private to you.
We also want to help those with speech disorders, or people whose speech has been affected by a stroke or ALS. Researchers from Google AI are exploring the idea of personalized communication models that can better understand different types of speech, as well as how AI can help even those who cannot speak to communicate. We call this research Project Euphonia.
Let’s take a look.
We are working hard to provide these voice recognition models through the Google Assistant in the future. But as you saw in Dimitri’s case, this will only be possible with many more speech samples to train our models on.
If you or someone you know has slurred or hard to understand speech, we’d like to invite you to submit voice samples to help accelerate this effort.
Fundamentally AI research which enables new products for people with disabilities is an important way we drive our mission forward. Live Transcribe, Live Caption, Live Relay, and Project Euphonia will ultimately result in products that work better for all of us. It’s a perfect example of what we mean by building a more helpful Google for everyone.
One of the most powerful ways we deliver help to our users is through our open source platforms like Android. To tell you more, I’d like to invite Steph onto the stage.
Stephanie Cuthbertson – Senior Director for Android
It’s amazing we’re here to talk about Android’s version 10. And we get to celebrate a milestone together. Today, there are over 2.5 billion active Android devices.
And today, we want to walk you through what’s coming next in the Android Q. Innovation, security, and privacy — the central theme of the Q release, and digital well-being.
A lot has changed since 1.0. Smartphones have evolved from an early vision to this integral tool in our lives. And they are incredibly helpful. Looking ahead, we see another big wave of innovation coming to make them even more helpful.
Q shows Android shaping the leading edge of mobile innovation, with over 180 device makers around the world. Driven by this powerful ecosystem, many innovations have been first on Android, from large screens to the first OLED display. And this year, display technology will take an even bigger leap with foldables coming from multiple Android OEMs.
These devices open up a completely new category, which, though early, just might change the future of mobile computing. Foldables take advantage of completely new display technology. They literally bend and fold from phone to tablet-sized screen. And Q maximizes what’s possible on these screens.
For instance, foldables are great for multitasking. So I can watch some funny videos my sister sent me while we chat about what we’re going to do for my mom on Mother’s Day. But the feature I’m most excited about is screen continuity. So let’s say we finish chatting. It’s time to head out. And I’m standing around, waiting for my ride.
So I start playing a game on the folded smaller screen. When I sit down and unfold, the game seamlessly transfers to the larger screen. It is so cool. And I can pick up exactly where I was playing. Now, multiple OEMs will launch foldables this year, all running Android.
Another exciting innovation is 5G. 5G networks mean consistently faster speeds with lower latency. So apps and especially games can target rich, immersive experiences to these 5G-connected phones. And Android Q supports 5G natively. This year, more than 20 carriers will launch networks. And our OEMs have over a dozen 5G-ready phones all launching this year. And they’ll all be running Android.
Now, in addition to hardware innovation, we’re also seeing huge firsts in software, driven by advances in on-device machine learning. Sundar showed Live Caption. Now, I would really like you to see it in action and then take you under the hood.
Please welcome Trystan.
Trystan: Like many people, I watch videos without sound when I’m on the go. With captions, I can still keep up, even if I’m in a crowded space or I’m sitting in a meeting. So for me, they’re super helpful. But for almost 500 million people who are deaf or hard of hearing, captions are critical.
Today, loads of mobile content embeds audio, from video to voice messages and everything in between. Without captions, this content is nowhere near as accessible. Live Caption in Q takes audio and instantly turns it into text. Let’s take a look at this video my friend Heather sent me yesterday.
To turn it on, I open the volume rocker and tap the Live Caption button.
[Heather: Hey, cutie. Do you want to give your puppy a hug? Oh. Oh, I guess not. Puppy is walking away.]
So as you can see, these captions appear in real time, over a video that would normally never have captions. You can expand them, contract them, move them up and down. It’s a lot of fun.
But what makes this feature so incredible is that it’s entirely done on device. In fact, it doesn’t need to be connected to the internet at all. If we take a look, this entire demo I’ve done in airplane mode.
Stephanie Cuthbertson – Senior Director for Android
Thank you, Trystan.
So how is this possible? It’s because of a huge breakthrough in speech recognition that we made earlier this year. This once required streaming audio to the cloud to run a two-gigabyte model for processing. Now we can do that same processing on device, using a recurrent neural net, in just 80 megabytes. The live speech model is running on the phone. And no audio stream ever leaves it.
All this protects user privacy. And this is OS-wide, which means you get those captions in all your apps and in web content too.
Now, the same on-device machine learning powers another useful Q feature, which is Smart Reply. With Smart Reply, the OS helpfully suggests what you’ll type next. It’ll predict the text you’ll type — even emoji. And it’s a huge time-saver.
What’s really cool is, this works now for all messaging apps in Android. Like in Signal, you can see the OS providing these helpful suggestions. And Smart Reply can now even predict the actions that you’ll take. So say a friend sends you an address. And normally, you copy and paste that into Maps. That’s kind of a hassle. With Smart Reply, you just tap and it will open for you. Now, all this is saving you time.
On-device machine learning powers everything from these incredible breakthroughs like Live Caption to helpful everyday features like Smart Reply. And it does this with no user input ever leaving the phone, all of which protects user privacy.
Now there’s one more addition to Android Q that’s small, but you’ve been asking us about for a while, and that is dark theme. And we’re launching it in Q. So you can activate it by using the quick tile or by turning on Battery Saver. And in fact, it will help you save battery. Your OLED display is one of the most power-hungry components in your phone. So by lighting up less pixels, it will save you battery. So that’s innovation.
But we feel all innovation must happen within a frame of security and privacy. People now carry phones constantly. And we trust them with a lot of personal information. You should always be in control of what you share and who you share it with. And that’s why the second area we’ll cover in the central focus of the release is security and privacy.
Now, over the years, Android’s built out a huge set of protections already — file-based encryption, SSL by default, secure DNS, work profiles. And many of these were first on Android. Android has the most widely deployed security and anti-malware service of any OS, with Google Play Protect. It runs on every device. And it scans over 50 billion apps a day.
In fact, in Gartner’s 2019 security report, which was published this week, Android scored the highest possible rating in 26 out of 30 categories. It’s ahead on multiple points, from authentication, to network security, to malware protection, and more. At the same time, we wanted to go much further. And that’s why Android Q includes almost 50 features focused on security and privacy, all providing more protection, transparency, and control.
So first, in Q, we brought privacy to the top level in settings. And there, you’ll find a number of important controls all in one place — activity data, location history, ad settings. And you decide what’s on or off.
Now, location is another place we’ve created tools for more transparency and control. Now, location can be really helpful, especially when you’re lost in a new place. But it’s also some of your most personal information. And you should, again, always be in control of who you share it with and how they can use it.
So first, if you’re wondering which apps can be accessing your location, we make it easy for you to know. With Q, your device will give you helpful reminders whenever an app accesses location when you’re not actively using that app. So you can review and decide, do you want to continue sharing or not?
Second, Q will give you more control over how you share location data with apps. For example, say you want to get pizza delivered. You can choose to share your location only while the app is in use. And as soon as you close, you’ll stop sharing location.
Finally, what if you’re wondering, what kind of location do all my apps have? In Q, we’ve brought location controls to the forefront in settings. So you can quickly review every app and change location access with simple controls.
Now, there are many, many more enhancements to security and privacy throughout the OS, like TLC v3, encryption for low-end devices, randomizing your Mac address by default, and many more. And you can read about all of these in our blog post this week.
But there’s one more really big thing for security. Now, your Android device gets regular security updates already. But you still have to wait for the release. And you have to reboot when they come. We want you to get these faster — even faster. And that’s why, in Q, we’re making a set of OS modules updatable, directly over the air. So now, these can be updated individually as soon as they are available, and without a reboot of the device.
Now, this was a huge technical challenge. We’re updating these in the background the same way we’re updating Google Apps. It’s easier for our partners, with whom we’re working closely. But more importantly, it’s much better for you. You can learn more about this at the session “What’s New in Android?“
Now, there’s one more thing that’s changed since the early days of Android. Now, people carry smartphones everywhere, because they’re really helpful. But we’re also spending a lot of time on phones. And people tell us sometimes they wish they’d spent more time on other things. We want to help people find balance and digital well-being.
And yes, sometimes, this means making it easier to put your device away entirely and focus on the times that really matter. That’s why, last year, we launched digital well-being tools with dashboards, app timers, Flip to Shush, and Wind Down to help you set the phone down and get to sleep at night. And these tools are really helping.
App timers help users stick to their goals over 90% of the time. And users of Wind Down had a 27% drop in nighttime usage. If you’re not using these already, I would really recommend them.
But this year, we want to help even more with distraction. A lot of times, I just want to sit down and focus to get something done. And when I’m trying to do this — like, working. Maybe it’s studying for you — I don’t want email or anything else to distract me. And that’s why we’ve created a new mode for Android. It’s called Focus mode.
When I enter Focus mode, I can select the apps that I find distracting. For me, that’s email, news. So now they’re turned off and I can really get to work. Those apps that distract me are disabled. But I can still keep text. Because it’s important to me that my family can always get a hold of me, until I come out of Focus mode. And then everything is back. Focus mode is coming to devices on P and Q this fall.
Now, finally, I want to talk about families. For 84% of us parents, technology use by our kids is a top concern. In the U.S., the average age of kids getting phones is now eight. In Q, Family Link parental controls will be built right into the settings of the device. So when you set up a device for someone in your family, Family Link will help connect it to a parent. And you can review any apps that your child wants to install.
After that, you can set daily screen time limits. You can check, where are the apps where my kids are spending time? And you can set a device bedtime so your kids can disconnect and get to sleep. And now on Android Q, you can set time limits on specific apps. And when your child hits that device bedtime, if you want to give them just five more minutes, now we have bonus time.
Now, there’s a ton more in Q that we don’t have time to cover — a ton — everything from streaming media, to hearing aids, to better connectivity, to new gesture UI, and more. So today, I’m excited to announce that Q Beta 3 is available on 21 devices. That is 12 OEMS plus all Pixels. And that is more than double last year.
We hope you head over to the link to get it on your phone, because we would love to have you try it out. And now I will hand it over to Rick.
Thank you very much.
Rick Osterloh – SVP of Hardware
Well, we’ve heard about some terrific innovations today in Android, AI, and the Assistant, and real breakthroughs in how we’re able to help our users.
I’d like to spend a few minutes and talk about how some of those come to life in our made by Google products. Now, we continue to believe that the biggest breakthroughs are happening at the intersection of AI, software, and hardware, whether that’s a tensor processing unit, an entire data center, the phone in your hand, or a helpful smart display in your home.
NEST HUB MAX
Let’s start there. The smart home of today is fragmented and frustrating. To deliver real help in the home, you can’t start with technology. You have to start with people. And we’ve always worked to put people first and build technology around their needs. There’s no more important place to get this right than in the home.
Let’s take a look.
Your home is the most special place in your life. So we need to be thoughtful about the technology we create for it. By putting people first, we’re going beyond the idea of a smart home to create a truly helpful home.
Over the past year, we’ve brought the Nest and Google teams together to deliver on our vision of the helpful home. And today, we’re further simplifying things, bringing all of these products together under the Nest name. As a single team and a single product family, we’re following a set of guiding principles that reflect our commitment to putting people first.
Now, to start, we believe technology should be easy for everyone in the home to use, whether they’re five or 95. The helpful home should also be personal for everyone. With Google Assistant at the core, we can provide a personalized experience for the entire household, even in communal spaces.
And the tech in your home should work together for a single seamless experience across rooms and devices. Most importantly, the helpful home needs to respect your privacy. And today, we’re publishing privacy commitments for our home products that clearly explain how they work, the data we’re storing, and how it’s used.
Our vision for the helpful home is anchored in the Assistant. And as you heard from Scott, we’re continuing to get more helpful over time.
We want to make sure that you can get the help you need where you need it. Google Home Hub, which we’re renaming Nest Hub, was designed specifically to bring the helpfulness of the Assistant to any room in your house. Now, we’ve also been working on a new display that builds on the things that people love about Hub, but is designed for communal spaces in the home where the family gathers.
Introducing Nest Hub Max. It’s a new product that has a camera and a larger 10-inch display, which is perfect for the center of your helpful home. Hub Max pulls together your connected devices into a home view dashboard, where you can see your Nest cams, you can switch on lights, control your music, and adjust your thermostat.
Hub Max also supports Thread. So just like Nest Connect, it communicates directly with Thread-supported devices that need a low power connection, like door locks or motion sensors.
And we’ve designed Hub Max width an incredibly helpful camera. If you want to know what’s going on in your home, you can choose to use it like a Nest Cam. You can turn it on when you’re away from home. You can check on things right from the Nest app in your phone. And just like a Nest Cam, it’s easy to see your event history, enable home and away assist, and you also get a notification if the camera detects any motion or sees someone it doesn’t recognize in your home.
Now, video calling is easy too with Google Duo. The camera has a wide angle lens. And it automatically adjusts to keep you centered in the frame. You can chat with any iOS or Android device, or a PC with a Chrome browser. You can also use Duo to leave video messages for members of your household.
Hub Max is designed to give you full control over the camera. Nothing is streamed or recorded unless you intentionally enable it. And you’ll always know when the camera is on, with a green indicator light. You have multiple controls to disable camera features. And a physical switch on the back electrically disconnects the camera and the microphones. And you can see all these controls clearly on the display.
Hub Max is designed to be used by multiple people in your home and provide everyone with the help they need in a personalized way. Now to help with that, we’ve offered users the choice to enable voice match, so the Assistant can recognize your voice and respond directly to you.
But today, we’re also extending the options to personalize using the camera, with a feature we call Face Match. For each person in your family that chooses to turn it on, the Assistant guides you through a process of creating a face model, which is then encrypted and stored on the device.
Then, whenever you walk in front of the camera, Hub Max recognizes you and shows just your information and not anyone else’s. Face Match’s facial recognition technology is processed locally on the device using on-device machine learning, so the camera data never leaves the device.
And in the morning, I can walk into the kitchen and the Assistant knows to greet me with my calendar, my commuting details, the weather, and any other information I need to start my day.
And when I get home, Hub Max welcomes me home with any reminders that might be waiting for me. And the Assistant offers personalized recommendations for music and TV shows. And I can even see if anyone’s left me a video message.
One of my favorite things about Hub Max is that it’s a great digital photo frame. No matter what kind of day I’m having, nothing makes me feel better than seeing some of my favorite memories on this beautiful screen. And the Google Photos integration makes this whole process really simple. I can select my family and friends. And Hub Max displays the best photos of them from years ago or from earlier today.
And now, with a simple voice command, sharing my favorite shots is easier than ever. The big screen also makes Hub Max the kitchen TV you’ve always wanted to. Tell it what you want to watch. Or if you need help deciding, just ask the Assistant to pull up our new onscreen guide.
Hub Max can stream your favorite live shows and sports on YouTube TV. But unlike your kitchen TV, it can also teach you how to cook, see who’s at the front door, and play your music.
You’re also getting full stereo sound with a powerful rear-facing woofer. And now, when the volume’s up, instead of yelling at the Assistant to turn it down or pause the game, with the camera, it’s as simple as a gesture. You just raise your hand. And Hub Max uses on-device machine learning to instantly identify your gesture and pause your media.
Hub Max is a Google Assistant smart display that’s also a smart home controller, a TV for your kitchen, a great digital photo frame, an indoor camera, and it’s perfect for video calling. All this will be available on Nest Hub Max later this summer for just $229.
And today, we’re lowering the price of the original Nest Hub from $149 to $129. And we’re expanding its availability to 12 new markets and supporting nine new languages.
So whether you prefer a Hub with a camera or without one, we have a device that’ll help you in your home.
As I said earlier, there is a fundamental difference between a smart home and a helpful home. And we’re excited to unify all our products under the Nest brand to make the helpful home more real for more people.
All right. Next, I want to talk about Pixel. Yeah, thank you.
And I love talking about Pixel. I want to talk about our work to bring a more hopeful smartphone experience to more people. A core element of Google’s mission is to make technology more available and accessible for everyone.
And Sundar said it earlier. We need to ensure that technology benefits the many, not just the few. But there’s been a really troubling trend in the smartphone industry. To support the latest technologies, everyone’s high-end phones are getting more and more expensive. So we challenged ourselves to see if we could optimize our software and AI to work great on more affordable hardware, so we can deliver these high-end experiences at a more accessible price point.
I want to introduce you to the newest members of the Pixel family, Google Pixel 3a and 3a XL, designed to deliver premium features at a price people will love. We didn’t compromise on the capabilities and performance you’d expect from a premium device, which is why we branded them Pixel. They start at just $399. That’s about half the price — half the price of typical flagship phones.
And I want to introduce Sabrina to tell you more about how we did it.
Sabrina Ellis – VP of Product Management
Delivering premium features with high performance on a phone at this price point — it’s been a huge engineering challenge. And I’m really proud of what our team has been able to accomplish with Pixel 3a.
So let’s start with the Design. Pixel 3a follows the design language of the Pixel family — the familiar two-tone look, smooth finish, and ergonomic unibody design. It feels good in your hand and it looks beautiful.
Pixel 3a comes in three colors, Just Black, Clearly White, and a new color, Purple-ish. Everything looks amazing on the vibrant OLED display. And your music, your podcasts, they sound great in premium stereo sound.
Pixel 3a supports Bluetooth 5.0 and USB-C digital audio. And we’ve also included a 3.5-millimeter audio jack. Because we’ve heard some people want more headphone options.
But what Pixel is really known for is its incredible camera. And with software optimizations, we found a way to bring our exclusive camera features and our industry-leading image quality into Pixel 3a, so photos look stunning in any light.
What other smartphone cameras try to do with expensive hardware, we can deliver with software and AI, including high-end computational photography. So here’s what that means. Pixel 3a can take amazing photos in low lights with Night Sight. It’s one of Pixel’s most popular features.
We’ve also enabled Pixel’s portrait mode on both the front and rear cameras. And our Super Res Zoom applies computational photography, so you can get closer to your subject while still maintaining a high degree of resolution. And all of your beautiful photos are backed up for free in high quality with Google Photos.
Pixel 3a also has the helpful features you’d expect in a Pixel. Just squeeze the sides of your phone to bring up the Google Assistant. We’re using the AI in Pixel 3a to help manage your phone calls too. I’m pretty sure we all hate getting robocalls. And Call Screen uses Google speech recognition and natural language processing to help you filter out those unwanted calls. It’s already screening millions of them.
Now, you might remember, last year, we shared our vision for using AR in Google Maps. Starting today on Pixel phones, when you use walking directions, instead of staring at that blue dot on your phone, you’re going to see arrows in the real world to tell you where to turn next. We’re just beginning our journey with AR in Maps. And we’re really excited for Pixel users to experience this early preview.
Battery life — it’s one of the most important features on a smartphone. It makes sense. People need to know that their phone won’t quit on them before the end of their day. Pixel 3a has adaptive battery. It uses machine learning to optimize based on how you use your phone, so you can get up to 30 hours on a single charge. And with the included 18-watt charger, you’ll get up to seven hours of battery life with just 15 minutes of charging.
Pixel 3a doesn’t compromise on Security either. It’s got the same comprehensive approach as Pixel 3. On the hardware side, our Titan M security chip protects your sensitive data on the device, like login credentials, disk encryption, app data, and OS integrity. On the software side, you get the latest Google security patches and updates for three years, including Android Q this summer.
So instead of getting slower and less secure over time, your Pixel gets better with every update. We think this hybrid approach provides the strongest data protection. And in a recent Gartner report, Pixel scored the highest for built-in security among smartphones.
Pixel 3a offers the complete Pixel experience. And we’re proud to make it available and affordable to more people around the world. Verizon’s been a great partner over the past two and a half years in the US. And we’re excited to be partnering with them again for the launch of Pixel 3a.
And for the first time, we’re expanding our US carrier partnerships. So the entire Pixel family is now available for sale at T-Mobile, Sprint, and US Cellular.
You can also get Pixel 3a from the Google Store and use it on any US carrier, including Google Fi and AT&T.
Pixel 3a and 3a XL are available in 13 markets, starting today. You can find more details online at the Google Store. We’re really excited to have you try it out.
Next, Jeff will tell you about our efforts in Google AI. But first, here’s a quick look at our new Pixel.
Jeff Dean – Lead of Google AI
Everything from building a low cost premium device like the one you just saw without compromising on capabilities to developing a truly helpful Assistant were all built on a tremendous amount of research and innovation under the covers. And they’re examples of what we do at Google AI.
Google AI is a collection of teams focused on making progress in artificial intelligence research across a wide range of different domains. We focus on solving fundamental computer science challenges in order to solve problems for people. That includes things like improving speech recognition models to answer questions faster and let you interact with your device quickly, or pushing the boundaries of computer vision to help people interact with their world in new ways, as you’ve seen today.
We publish papers, release open source software, and apply our research to Google products. The goal is really to solve problems everyday that touch billions of people.
One of the things I’m most excited about is progress in language understanding. As Scott mentioned earlier, so much of our daily life depends on actually understanding language — reading traffic signs and shopping lists, writing emails, communicating with the people around us.
We’d really want computers to have the same fluency with language that we do — not just understand surface forms of the words, but actually understand what sentences mean. Unlocking that would get us closer to our mission of organizing the world’s information and making it universally accessible and useful.
In the past few years, we’ve made major strides. Take teaching a machine to answer questions like this one about Carlsbad Caverns, a national park in New Mexico. Only recently, the state-of-the-art architecture for language understanding was something called a recurrent neural network, or RNN.
RNNs process words sequentially, one after another. They work well for modeling short sequences, like sentences, but they struggle to make abstract associations, like knowing that stalactites and stalagmites are natural formations, and that cement pathways, for example, are not.
In 2017, we made a leap forward with our research on transformers, models that process words in parallel. One year later, we used it as the foundation for a technique we called bi-directional encoder representations from transformers. It’s a bit of a mouthful, so we just call it BERT.
BERT models can consider the full context of a word by looking at the words that come before and after it. They’re pre-trained using plain text from the web and other textual sources. To do that, we use a process to train it that’s a little like the word game of Mad Libs. We hired about 20% of the input words. And we train the model to guess those missing words. You can actually try this at home with a bit of text that you have. Hide a few words and see if you can guess what they are. That’s effectively what we’re doing.
This approach is much more effective for understanding language. When we published the research, BERT obtained state-of-the-art results on 11 different language processing tasks.
Fast forward to today. And we’re excited to see how BERT can help us answer more complex questions that are relevant to you, whether that’s getting the flight time from Indiana to Honolulu, learning a new weightlifting exercise, or translating between languages. Research like this gets us closer to technology that can truly understand language.
We’re now working with product teams all across Google to see how we can use BERT to solve more problems in more places. We’re excited to bring this to people around the world to help them get the information they need everyday.
All this machine learning momentum, though, wouldn’t be possible without platform innovation. TensorFlow is the software infrastructure that underlies our work in machine learning and artificial intelligence. When we developed TensorFlow, we wanted everyone to be able to use machine learning, so we made it an open source platform.
And while it’s been essential to our work, we’ve been amazed to see what other people outside of Google have used it for — all kinds of different things. We’ve seen engineers at Roma Tre University in Italy parsing handwritten medieval manuscripts. We’ve seen coders in France colorizing black and white photography. We’ve even seen companies developing fitness centers for cows.
The work that people are doing is really inspiring to us. It pushes us to keep asking ourselves, how can machine learning crack open previously unsolvable problems in order to help more people?
One example is our work in the field of health care. We’re really optimistic that our research can create real world impact in medicine by improving solutions and establishing new diagnostic procedures. To share more, here’s Dr. Lily Peng from the Google AI health care team.
Lily Peng – Product Manager, Google AI Healthcare Team
So as a doctor, what I care about most is improving patients’ lives. And that means good care and accurate diagnoses. That’s why I was so excited two years ago at I/O when we shared our work in diabetic retinopathy. This is a complication of diabetes that puts over 400 million people around the world at risk for vision loss.
Since then, we’ve been piloting this work with patients in clinical settings. Our partners have fairly recently received European regulatory approval for the machine learning model. And we have clinical deployments in Thailand and India that are already screening thousands of patients.
In addition to diabetes, one of the other areas we think AI can help doctors is in oncology. Today, we’d like to share our work on another project in cancer screening where AI can help catch lung cancer earlier. So lung cancer causes more deaths than any cancer. It’s actually the most common cause of death globally, accounting for 3% of annual mortality.
We know that, when cases are diagnosed early, patients have a higher chance of survival. But unfortunately, over 80% of lung cancers are not caught early.
Randomized control trials have shown that screening with low dose CTs can help reduce mortality. But there’s opportunity to make them more accurate.
So in a paper we are about to publish in “Nature Medicine,” we describe a deep learning model that can analyze CT scans and predict lung malignancies. To do it, we trained a neural network with de-identified lung cancer scans from our partners at the NCI — the National Cancer Institute — and Northwestern University.
By looking at many examples, the model learns to detect malignancy with performance that meets or exceeds that of trained radiologists.
So concretely, how might this help? Very early stage cancer is minuscule, and can be hard to see, even for seasoned radiologists, which means that many patients with late stage lung cancer have subtle signs on earlier scans. So take this case, where an asymptomatic patient with no history of cancer had a CT scan for screening. This scan was interpreted as normal.
One year later, that same patient had another scan. It picked up a late stage cancer, one that’s much harder to treat. So we used our AI system to review that initial scan. So let’s be clear. This is a tough case. We showed this initial scan to other radiologists, and five out of six missed this cancer.
But our model was able to detect these early signs one year before the patient was actually diagnosed — one year. And that year could translate to an increased survival rate of 40% for patients like this.
So clearly, this is a promising but early result. And we’re very much looking forward to partnering with the medical community to use technology like this to help improve outcomes for patients.
Now I’ll hand it back to Jeff.
Jeff Dean – Lead of Google AI
Thanks, Lily. The same technologies that you just saw driving health care innovation have applications across almost every field imaginable.
Our AI for Social Good program brings together our efforts to use AI to explore and address some of the world’s most challenging problems. Last year, we announced the program and its two pillars — research and engineering, and building the external ecosystem.
Let’s talk first about Research & Engineering. One project we’re working that’s already creating impact is our work on flood forecasting. Floods are the most common, deadliest natural disasters on the planet. Every year, they affect up to 230 million people across the world, more than storms and earthquakes combined.
20% of flood fatalities happen in India alone. This is a problem that we’re even seeing this week with the impact from Cyclone Fani. Floods prevent kids from being able to play in their neighborhoods or parents from protecting and providing for their families, often because they don’t have enough advance warning.
And without consistent accurate warning systems, people are prone to ignore warnings and be unprepared. That’s especially detrimental in areas hit with annual monsoons. That’s why, last fall, we shared our work on flood forecasting models that can more accurately predict flood timing, location, and severity.
Through a partnership with India’s Central Water Commission, we began sending early flood warnings to the phones of users who might be affected. Today, we’re thrilled to announce the expansion of our detection and alerting system for the upcoming monsoon season. The expanded area will cover millions of people living along the Ganges and Brahmaputra River areas.
Not only are we increasing the area of coverage, but we’re also better forecasting where the floods will hit hardest. Through a new version of our public alerts, people can better understand whether they’ll be affected, so they can protect themselves and their families.
Our model simulates water behavior across the flood plain, showing the exact areas that will be affected. We combine thousands of satellite images to create high resolution elevation maps, using a process similar to stereographic imaging, to figure out the height of the ground.
We then use neural networks to correct the terrain, so it’s even more accurate. And then we use physics to simulate how flooding will happen. We also collaborate with the government to receive up-to-date stream gauge measurements and send forecasts in real time.
We’re excited to continue working with partners to increase the accuracy and precision of these models, which we hope will make people safer from flooding all around the world.
Research like this is critical. But we also know that AI will have the biggest impact when people from many different backgrounds all come together to develop new solutions to problems they see. That’s why the second pillar of our AI for Social Good program is to build the external ecosystem. We want to empower everyone to use AI to solve problems they see in their communities.
Last year, we partnered with Google.org to launch the Google AI Impact Challenge. It was a call for nonprofits, social enterprises, and universities to share their ideas for using AI to address societal challenges. We received applications from 119 countries across six continents, representing all kinds of sizes and types of organizations.
Today, we’re really excited to announce the 20 selected. We even have a few of them with us today. Let’s give them a warm welcome.
Here’s the list of organizations. These organizations are working on some of the world’s most meaningful issues. La Fondation Medicins sans Frontieres is using image recognition to help medical staff analyze anti-microbial images in order to prescribe the right antibiotics for bacterial infections.
New York University in partnership with the Fire Department of New York City is building a model to help speed up emergency response times. This could really improve public health and safety.
And Makerere University in Uganda will use AI to create a high resolution monitoring network to shape public policies for improving air quality. We’ll be supporting our 20 grantees and bringing these ideas to life. We’re providing $25 million in funding from Google.org, as well as coaching and resources from teams all across Google.
Congratulations to all our grantees.
As we head into the next decade, I’m really excited about what’s to come. There are so many promising avenues for fundamental research. For instance, machine learning models today, typically, we can get them to be good at solving individual tasks.
But what if they could generalize across thousands of tasks, solving new problems faster, and with just a few examples to learn from? The keys to progress on these kinds of research problems are those most human characteristics, perseverance and ingenuity.
As you heard Sundar mention at the start of the day, we’re moving from a company that helps you find answers to a company that also helps you get things done. And all the products we showed you today share a single goal — to be helpful.
At the same time, we want to ensure that the benefits of technology are felt everywhere, continue to uphold our foundation of user trust, and build a more helpful Google for everyone.
To everyone joining us on the livestream, thank you for tuning in. And to everyone here with us in the audience today, welcome to Google I/O 2019.
Thank you. And enjoy the rest of I/O.
- The Battle for Your Time: Exposing the Costs of Social Media: Dino Ambrosi (Transcript)
- TRANSCRIPT: The Powerful Possibilities of Recycling the World’s Batteries: Emma Nehrenheim
- The Exciting, Perilous Journey Toward AGI: Ilya Sutskever (Transcript)
- Why AI Is Incredibly Smart and Shockingly Stupid: Yejin Choi (Transcript)
- The Truth About Mobile Phone And Wireless Radiation: Dr Devra Davis (Transcript)