Tag Archives: AI

Google Research India: an AI lab in Bangalore

People have been using technology to solve problems and improve their quality of life for centuries, from sharing knowledge with the printing press to going online to build a small business. These days, artificial intelligence is opening up the next phase of technological advances. And with its world-class engineering talent, strong computer science programs and entrepreneurial drive, India has the potential to lead the way in using AI to tackle big challenges. In fact, there are already many examples of this happening in India today: from detecting diabetic eye disease to improving flood forecasting and teaching kids to read


To take this to the next level we've created Google Research India—an AI lab we’re starting in Bangalore. This team will focus on two pillars: First, advancing fundamental computer science and AI research by building a strong team and partnering with the research community across the country. Second, applying this research to tackle big problems in fields like healthcare, agriculture, and education while also using it to make apps and services used by billions of people more helpful. 


Google Research India will be led by Manish Gupta, a renowned computer scientist and ACM Fellow with a background in deep learning across video analysis and education, compilers and computer systems. We’re also excited to have Professor Milind Tambe join us on a joint appointment from Harvard University as Director of AI for Social Good. Professor Tambe will build a research program around applying AI to tackle big problems in areas like healthcare, agriculture, or education. 


The lab in Bangalore will be part of and support Google’s global network of researchers: participating in conferences, publishing research in scientific papers, and collaborating closely with one another. We’re also exploring the potential for partnering with India’s scientific research community and academic institutions to help train top talent and support collaborative programs, tools and resources. 


Starting Google Research India is an important step for us, and for me personally, too. As someone who grew up in India, studied at the Indian Institute of Science in Bangalore, and learned so much from the community there, I’m grateful that we now have the opportunity to help advance research and play a part in building the AI community in India. 


If you’re a scientist or researcher interested in learning more about Google Research India, click here.

Tracking our progress on flood forecasting

Around the world, floods cause between 6,000 and 18,000 fatalities every year—20 percent of those are in India—and between $21 and $33 billion in economic damages. Reliable early warning systems have been shown to prevent a significant fraction of fatalities and economic damage, but many people don’t have access to those types of warning systems.


Last year we began our flood forecasting pilot initiative in the Patna region of India, with the goal of providing accurate real-time flood forecasting information and alerts to those in affected regions. This is made possible through AI and physics-based modeling, which incorporate data from historical flooding events, river levels, terrain and elevation data. We generate high-resolution elevation maps and run up to hundreds of thousands of simulations in each location. With information obtained through our collaboration with Indian Central Water Commission, we create river flood forecasting models that can more accurately predict not only when and where a flood might occur, but the severity of the event as well. Here’s a bit more on the progress we’ve made over the past year.

Expanding our coverage area and tracking accuracy 

Our coverage area is now twelve times greater than it was last year—covering 11,600 sq. kilometers in India along the Ganga and Brahmaputra rivers, two of the most flood affected rivers in the world. We’ve also sent over 800,000 alerts to individuals in affected areas. 

map.png

Coverage area of our current operational flood forecasting systems.

The alerts we send out include three tiers of risk (covering approximately equal areas): Some flood risk, greater flood risk, and greatest flood risk.

flood map.png

Accuracy and reliability are paramount to the success of the initiative, and the safety of those in affected areas. Incorrect forecasts do more harm than good, and vague or overly-general warnings are consistently ignored by affected populations. We track the accuracy of our alerts across two main metrics:

Some-risk recall

This tells us what percentage of the actual flood was covered by our “some risk” warning. If this metric is low, that means people who should be warned are not getting warned. During this monsoon season, our some-risk recall metric was well over 95 percent, which means the vast majority of affected areas were correctly forecasted to be flooded.

Greatest-risk precision

This tells us what percentage of our greatest-risk forecast ended up being flooded. If this metric is low, that means we’ve told a lot of people we’re confident they’re at risk while they weren’t, and we may lose people’s trust. During this monsoon season, our high-risk precision was around 80 percent, which means that people who received a high-risk warning were indeed very likely to be affected.

To enable even faster progress in the future, we've increased the efficiency of our simulation models, automating manual tasks, and experimenting with new forecast methodologies. Read more about how we do this on the Google AI blog.

flood simulation.png

A snapshot of a TPU-based simulation of flooding in Goalpara, mid-event.

Distributing alerts

Accurate and reliable flood forecasting can help keep people safe, if they’re getting the early warning. Over the past year, we’ve significantly improved our notification infrastructure. We’ve expanded the products we use to inform affected individuals, by adding better crisis information on Google Maps. Whether they find out about the flooding event through Search, Maps, or an alert push notification, people can quickly access an interactive map where they can see their location relative to where the flood is predicted to be.
PublicAlerts_v03.gif

Flood-related information

Key partnerships in local communities

Many people don’t have access to the internet, especially during a crisis, so we’ve partnered with the Indian nonprofit SEEDS to provide information to individuals on the ground. We provide SEEDS with our real-time forecasting information, and they directly interact with the community and local panchayats (village councils). This year we’ve piloted this system in Patna, sending alerts to local leaders across hundreds of villages in the district, and we’re actively collecting feedback about how well the system worked.

Of course, our most important partner is the government itself, which not only provides us with critical real-time data, but is also best-placed to provide relief efforts in times of crises. In the past year, we’ve developed a partner notification infrastructure to provide our forecasts for the Central Water Commission and other organizational partners, and are continuously working with them to improve this system to be more useful for disaster management efforts.

Flood management is an enormous challenge, and reducing the immense harms of floods globally will require a collaboration between governments, international organizations, the academic community, and operational experts. By continuing this work, we hope to help develop tools to make forecasts and response systems more accurate and accessible to everyone.


2,602 uses of AI for social good, and what we learned from them

For the past few years, we’ve applied core Google AI research and engineering to projects with positive societal impact, including forecasting floods, protecting whales and predicting famine. Artificial intelligence has incredible potential to address big social, humanitarian and environmental problems, but in order to achieve this potential, it needs to be accessible to organizations already making strides in these areas. So, the Google AI Impact Challenge, which kicked off in October 2018, was our open call to organizations around the world to submit their ideas for how they could use AI to help address societal challenges.


Accelerating Insights from the Google AI Impact Challenge” sheds light on the range of organizations using AI to address big problems. It also identifies several trends around the opportunities and challenges related to using AI for social good. Here are some of the things  we learned—check out the report for more details.


AI is globally relevant 

We received 2,602 applications from six continents and 119 countries, with projects addressing a wide range of issue areas, from education to the environment. Some of the applicants had experience with AI, but 55 percent of not-for-profit organizations and 40 percent of for-profit social enterprises reported no prior experience with AI. 


Goog


Similar projects can benefit from shared resources

When we reviewed all the applications, we saw that many people are trying to tackle the same problems and are even using the same approaches to do so. For example, we received more than 30 applications proposing to use AI to identify and manage agricultural pests. The report includes a list of common project submissions, which will hopefully encourage people to collaborate and share resources with others working to solve similar problems.  


You don’t need to be an expert to use AI for social good

AI is becoming more accessible as new machine learning libraries and other open-source tools, such as Tensorflow and ML Kit, reduce the technical expertise required to implement AI. Organizations no longer need someone with a deep background in AI, and they don’t have to start from scratch. More than 70 percent of submissions, across all sectors and organization types, used existing AI frameworks to tackle their proposed challenge. 


Successful projects combine technical ability with sector expertise 

Few organizations had both the social sector and AI technical expertise to successfully design and implement their projects from start to finish. The most comprehensive applications established partnerships between nonprofits with deep sector expertise, and academic institutions or technology companies with technical experience.


ML isn’t the only answer 

Some problems can be addressed by using alternative methods to AI—and result in faster, simpler and cheaper execution. For example, several organizations proposed using machine learning to match underserved populations to legal knowledge and tools. While AI could be helpful, similar results could be achieved through a well-designed website. While we’ve seen the impact AI can have in solving big problems, you shouldn’t rule out more simple approaches as well. 


Global momentum around AI for social good is growing—and many organizations are already using AI to address a wide array of societal challenges. As more social sector organizations recognize AI’s potential, we all have a role to play in supporting their work for a better world. 


Learning Cross-Modal Temporal Representations from Unlabeled Videos



While people can easily recognize what activities are taking place in videos and anticipate what events may happen next, it is much more difficult for machines. Yet, increasingly, it is important for machines to understand the contents and dynamics of videos for applications, such as temporal localization, action detection and navigation for self-driving cars. In order to train neural networks to perform such tasks, it is common to use supervised training, in which the training data consists of videos that have been meticulously labeled by people on a frame-by-frame basis. Such annotations are hard to acquire at scale. Consequently, there is much interest in self-supervised learning, in which models are trained on various proxy tasks, and the supervision of those tasks naturally resides in the data itself.

In “VideoBERT: A Joint Model for Video and Language Representation Learning” (VideoBERT) and “Contrastive Bidirectional Transformer for Temporal Representation Learning” (CBT), we propose to learn temporal representations from unlabeled videos. The goal is to discover high-level semantic features that correspond to actions and events that unfold over longer time scales. To accomplish this, we exploit the key insight that human language has evolved words to describe high-level objects and events. In videos, speech tends to be temporally aligned with the visual signals, and can be extracted by using off-the-shelf automatic speech recognition (ASR) systems, and thus provides a natural source of self-supervision. Our model is an example of cross-modal learning, as it jointly utilizes the signals from visual and audio (speech) modalities during training.
Image frames and human speech from the same video locations are often semantically aligned. The alignment is non-exhaustive and sometimes noisy, which we hope to mitigate by pretraining on larger datasets. For the left example, the ASR output is, “Keep rolling tight and squeeze the air out to its side and you can kind of pull a little bit.”, where the actions are captured by speech but the objects are not. For the right example, the ASR output is, “This is where you need to be patient patient patient,” which is not related to the visual content at all.
A BERT Model for Videos
The first step of representation learning is to define a proxy task that leads the model to learn temporal dynamics and cross-modal semantic correspondence from long, unlabeled videos. To this end, we generalize the Bidirectional Encoder Representations from Transformers (BERT) model. The BERT model has shown state-of-the-art performance on various natural language processing tasks, by applying the Transformer architecture to encode long sequences, and pretraining on a corpus containing a large amount of text. BERT uses the cloze test as its proxy task, in which the BERT model is forced to predict missing words from context bidirectionally, instead of just predicting the next word in a sequence.

To do this, we generalize the BERT training objective, using image frames combined with the ASR sentence output at the same locations to compose cross-modal “sentences”. The image frames are converted into visual tokens with durations of 1.5 seconds, based on visual feature similarities. They are then concatenated with the ASR word tokens. We train the VideoBERT model to fill out the missing tokens from the visual-text sentences. Our hypothesis, which our experiments support, is that by pretraining on this proxy task, the model learns to reason about longer-range temporal dynamics (visual cloze) and high-level semantics (visual-text cloze).
Illustration of VideoBERT in the context of a video and text masked token prediction, or cloze, task. Bottom: visual and text (ASR) tokens from the same locations of videos are concatenated to form the inputs to VideoBERT. Some visual and text tokens are masked out. Middle: VideoBERT applies the Transformer architecture to jointly encode bidirectional visual-text context. Yellow and pink boxes correspond to the input and output embeddings, respectively. Top: the training objective is to recover the correct tokens for the masked locations.
Inspecting the VideoBERT Model
We trained VideoBERT on over one million instructional videos, such as cooking, gardening and vehicle repair. Once trained, one can inspect what the VideoBERT model learns on a number of tasks to verify that the output accurately reflects the video content. For example, text-to-video prediction can be used to automatically generate a set of instructions (such as a recipe) from video, yielding video segments (tokens) that reflect what is described at each step. In addition, video-to-video prediction can be used to visualize possible future content based on an initial video token.
Qualitative results from VideoBERT, pretrained on cooking videos. Top: Given some recipe text, we generate a sequence of visual tokens. Bottom: Given a visual token, we show the top three future tokens forecast by VideoBERT at different time scales. In this case, the model predicts that a bowl of flour and cocoa powder may be baked in an oven, and may become a brownie or cupcake. We visualize the visual tokens using the images from the training set closest to the tokens in feature space.
To verify if VideoBERT learns semantic correspondences between videos and text, we tested its “zero-shot” classification accuracy on a cooking video dataset in which neither the videos nor annotations were used during pre-training. To perform classification, the video tokens were concatenated with a template sentence “now let me show you how to [MASK] the [MASK]” and the predicted verb and noun tokens were extracted. The VideoBERT model matched the top-5 accuracy of a fully-supervised baseline, indicating that the model is able to perform competitively in this “zero-shot” setting.

Transfer Learning with Contrastive Bidirectional Transformers
While VideoBERT showed impressive results in learning how to automatically label and predict video content, we noticed that the visual tokens used by VideoBERT can lose fine-grained visual information, such as smaller objects and subtle motions. To explore this, we propose the Contrastive Bidirectional Transformers (CBT) model which removes this tokenization step, and further evaluated the quality of learned representations by transfer learning on downstream tasks. CBT applies a different loss function, the contrastive loss, in order to maximize the mutual information between the masked positions and the rest of cross-modal sentences. We evaluated the learned representations for a diverse set of tasks (e.g., action segmentation, action anticipation and video captioning) and on various video datasets. The CBT approach outperforms previous state-of-the-art by significant margins on most benchmarks. We observe that: (1) the cross-modal objective is important for transfer learning performance; (2) a bigger and more diverse pre-training set leads to better representations; (3) compared with baseline methods such as average pooling or LSTMs, the CBT model is much better at utilizing long temporal context.
Action anticipation accuracy with the CBT approach from untrimmed videos with 200 activity classes. We compare with AvgPool and LSTM, and report performance when the observation time is 15, 30, 45 and 72 seconds.
Conclusion & future work
Our results demonstrate the power of the BERT model for learning visual-linguistic and visual representations from unlabeled videos. We find that our models are not only useful for zero-shot action classification and recipe generation, but the learned temporal representations also transfer well to various downstream tasks, such as action anticipation. Future work includes learning low-level visual features jointly with long-term temporal representations, which enables better adaptation to the video context. Furthermore, we plan to expand the number of pre-training videos to be larger and more diverse.

Acknowledgements
The core team includes Chen Sun, Fabien Baradel, Austin Myers, Carl Vondrick, Kevin Murphy and Cordelia Schmid. We would like to thank Jack Hessel, Bo Pang, Radu Soricut, Baris Sumengen, Zhenhai Zhu, and the BERT team for sharing amazing tools that greatly facilitated our experiments. We also thank Justin Gilmer, Abhishek Kumar, Ben Poole, David Ross, and Rahul Sukthankar for helpful discussions.

Source: Google AI Blog


Making learning to read accessible and fun with Bolo

The ability to read builds the foundation for education and a better life. Yet according to the United Nations, 617 million children and adolescents worldwide are not acquiring basic literacy skills. This number includes children like 9-year-old Alifya from Wazirpur, Delhi. Her parents, Shabana and Ramirrudin, understand the importance of an education and send her to school, but simply cannot afford to buy books or teach her themselves. As a result, her parents tell us Alifya is already behind her reading grade level and increasingly finding it hard to stay engaged. 

At Google, we believe technology can help kids around the world, like Alifya, learn how to read and can move us closer to the goal of basic universal literacy. Over the years, we've invested in this goal through our products, partnerships and funding. Google.org granted $50M and technical expertise to nonprofit innovators using technology to close education gaps. These organizations have reached more than 30 million students and are improving student outcomes and teacher effectiveness. They are ensuring technology improves everyone’s education experience, no matter their location or learning environment. And given the advancements in speech technology and AI, we believe there is room to do more.  

Earlier this year, we took one step forward in this direction and released Bolo, an AI-enabled Android app to help kids improve reading skills.

A mother and child using Bolo

Alifya’s mom helping her use Bolo.

We designed Bolo to act as a personal reading tutor to help any student whenever and wherever they need it. It uses speech-based technology to provide personalized assistance in a student’s reading journey, correcting them when they need help and encouraging them when they get it right. The app even works completely offline on low cost phones, which means children who need it the most also have access to the app.

When we tested Bolo in India, we found that 64 percent of kids who used the app showed an improvement in reading proficiency. In the five months since the launch in India, more than 800,000 children have used the app to read more than three million stories. “Earlier I couldn’t check Alfiya’s reading abilities. However, with Bolo, I can see her progress at home when she is using the app,” Shabana, Alfiya’s mother, tells us. “Her reading skills have improved and I believe this has increased her confidence.” These results have inspired us to expand our efforts to even more children in more places. 

Later this month, Bolo will also support Spanish and Portuguese, in addition to Hindi and English. This expansion will help children learn how to read in four of the most spoken languages in the world.  

Bolo in Spanish

As we hear more from students like Alfiya, who was able to significantly improve her reading skills after just three months, we are optimistic that technology can play a role in improving literacy. After all, every child who learns how to read is another student empowered to become a future author, doctor, artist, computer scientist or, in Alfiya's case, a teacher. 

MLIR: accelerating AI with open-source infrastructure

Machine learning now runs on everything from cloud infrastructure containing GPUs and TPUs, to mobile phones, to even the smallest hardware like microcontrollers that power smart devices. The combination of advancements in hardware and open-source software frameworks like TensorFlow is making all of the incredible AI applications we’re seeing today possible--whether it’s predicting extreme weather, helping people with speech impairments communicate better, or assisting farmers to detect plant diseases


But with all this progress happening so quickly, the industry is struggling to keep up with making different machine learning software frameworks work with a diverse and growing set of hardware. The machine learning ecosystem is dependent on many different technologies with varying levels of complexity that often don't work well together. The burden of managing this complexity falls on researchers, enterprises and developers. By slowing the pace at which new machine learning-driven products can go from research to reality, this complexity ultimately affects our ability to solve challenging, real-world problems. 


Earlier this year we announced MLIR, open source machine learning compiler infrastructure that addresses the complexity caused by growing software and hardware fragmentation and makes it easier to build AI applications. It offers new infrastructure and a design philosophy that enables machine learning models to be consistently represented and executed on any type of hardware. And today we’re announcing that we’re contributing MLIR to the nonprofit LLVM Foundation. This will enable even faster adoption of MLIR by the industry as a whole.
MLIR industry partners

MLIR aims to be the new standard in ML infrastructure and comes with strong support from global hardware and software partners including AMD, ARM, Cerebras, Graphcore, Habana, IBM, Intel, Mediatek, NVIDIA, Qualcomm Technologies, Inc, SambaNova Systems, Samsung, , Xiaomi, Xilinx—making up more than 95 percent of the world’s data-center accelerator hardware, more than 4 billion mobile phones and countless IoT devices. At Google, MLIR is being incorporated and used across all our server and mobile hardware efforts.


Machine learning has come a long way, but it's still incredibly early. With MLIR, AI will advance faster by empowering researchers to train and deploy models at larger scale, with more consistency, velocity and simplicity on different hardware. These innovations can then quickly make their way into products that you use every day and run smoothly on all the devices you have—ultimately leading to AI being more helpful and more useful to everyone on the planet.

Tackling cardiovascular disease with AI

Westmead team with Google’s Mel Silva and Australian Minister for Industry, Science and Technology, Hon Karen Andrews MP


Heart disease and cardiovascular health are a major challenge around the world, and in Australia, one in six people is affected by cardiovascular disease. The University of Sydney’s Westmead Applied Research Centre is working on a digital health program for people at risk of cardiovascular disease, and they recently received a $1 million Google.org grant that will help them apply AI to give patients more personalised advice and support.  

We sat down with Professor Clara Chow, Professor of Medicine and Academic Director at Westmead Applied Research Centre, and Dr. Harry Klimis, a cardiologist and Westmead PhD student, to hear more about the program.   

Why is cardiac health such a big issue? 

Professor Chow: Cardiovascular disease is the leading cause of premature death and disability worldwide. In Australia, cardiovascular disease affects approximately 4.2 million people, has resulted in more than 1 million hospitalizations, and caused 1 in 3 deaths in 2016. That’s one death every 12 minutes, and these deaths are largely preventable.

How are you proposing to address this problem? 

Chow: Our goal is to support people at high risk of developing cardiovascular disease by encouraging them to adopt healthy habits, such as diet and exercise, and connecting them to health services when they need them. Data and mobile technology means we can do this in ways that weren’t possible before. 

Dr Klimis: We’ve already developed mobile health text-message programs using basic algorithms to customise programs to individuals. We now plan to use machine learning and AI to keep improving how we support participants and help them self-monitor measures like cholesterol, blood pressure, weight, physical activity, diet and smoking.

How will you use the funding and support from Google.org? 

Chow: The grant will help us create digital tools that enable clinicians and health services to provide personalized advice without the need to meet face to face. Initially, we’ll link data from existing secondary sources like hospital and clinic presentations to create programs tailored to individuals, and the system will learn from there. 

How does AI help?  

Klimis: An example would be if “John” went to the emergency room at hospital with chest pain and had type 2 diabetes, obesity and hypertension. After being assessed and treated, he could be flagged as a patient at high risk of heart attack and added to the mobile health prevention program. The AI program would learn from John’s activities and deliver health advice via SMS or through an app. If John was less active at a particular time of day, the program might register this and prompt him to take a 5-minute walk. 

What do you think is going to be the most challenging part of your project?

Klimis: Making sure we have reliable enough data to support a program capable of AI and machine learning. Our original program sent out standard text messages to over 3000 people, which allowed us—with their permission—to collect data on their characteristics, how they respond to different messages, and how this affects health outcomes. That data will be crucial in building an AI model for the current project.  

What are you most optimistic about?

Chow: We have the potential to help more people at risk of cardiovascular disease by giving them high-quality prevention programs developed by clinicians and researchers, without requiring frequent clinic or hospital visits. Over the long term, mobile and digital health solutions could reduce hospitalizations, bring down healthcare costs, and make healthcare more accessible.  


When students get stuck, Socratic can help

In building educational resources for teachers and students, we’ve spent a lot of time talking to them about challenges they face and how we can help. We’ve heard that students often get “stuck” while studying. When they have questions in the classroom, a teacher can quickly clarify—but it’s frustrating for students who spend hours trying to find answers while studying on their own. 

Socratic, a mobile learning app we acquired last year, now uses AI technology to help high school and university students when they’re doing school work outside the classroom. It guides them through the resources that are available, and identifies the core underlying concepts that will lead them to answers to their questions.

With the latest improvements, here are a few ways Socratic has become even more helpful for students.

Get help at any moment

Students can take a photo of a question or use their voice to ask a question, and we’ll find the most relevant resources from across the web.  If they’re struggling to understand textbook content or handouts, they can take a picture of the page and check out alternative explanations of the same concepts.

Solving a math equation and a physics problem with Socratic’s help

 Solving a math equation and a physics problem with Socratic’s help


Understand the underlying concepts

To help students working on complex problems, we’ve built and trained algorithms that look at a student’s question and automatically identify the relevant underlying concepts. From there, we can find the videos, concept explanations, and online resources to help them work through their questions. For students who want to learn even more, we break down the concepts into smaller, easy-to-understand lessons. 

Socratic takes a problem, X-rays it, and extracts the underlying concepts.

Socratic takes a problem, X-rays it, and extracts the underlying concepts.

Browse helpful topic explanations for quick reference

To help students who are reviewing what they’ve learned or studying for a test, we’ve worked with educators to create subject guides on over 1,000 higher education and high school topics. It takes two taps to look up any topic they need to brush up on, get the key points, and then go deeper with helpful resources on the web.

Scroll on the Socratic app to find study guides and resources

Scroll on the Socratic app to find study guides and resources

You can leave feedback in the app- we’d love to hear from you. It’s available today on iOS and will be available on Android in the fall. 

An environmental nonprofit takes on AI “sprint week”

This May, the global group of Google AI Impact Challenge grantees gathered in San Francisco to kick off the six-month Launchpad Accelerator program. With $25 million in funding from Google.org, credits from Google Cloud and mentorship by Google’s AI experts, the teams sought to apply AI to address a wide range of problems problems, from protecting rainforests to coaching students on writing skills. 

Now in the second phase of the program, Tech Sprint Week, the grantees tackled their projects’ greatest technical challenges with support from a team of mentors from Google. At Google for Startups’ campus in London, teams continued work on their ideas and learned user experience design principles along the way.

Grace Mitchell, a data scientist at grantee WattTime, opened up about her team’s experience at Tech Sprint Week—and how they’re using AI to build a globally accessible, open-source fossil fuel emissions monitoring platform for power plants.

Can you tell us about WattTime? 

WattTime is an environmental tech nonprofit, and our mission statement is to give people the power to choose clean energy. Users integrate our API into their IOT (Internet of Things) capable devices, which tells them the type of fuel that provides their energy. It also tells them the environmental impact of the type of fuel they’re using. As an example, coal has a value equivalent to around 900 to 1200 pounds of emissions per megawatt hour, whereas renewable energy would be zero.The whole point is to shift electricity usage based on high or low emission periods. 

For this program, we’ve partnered with The Carbon Tracker Initiative to take on a new challenge: fossil fuel emissions monitoring. We’re using image processing algorithms and satellite networks to replace expensive, on-site power plant emissions monitors with a globally accessible, open-source monitoring platform.

Who is on your team for this project?

Our project for the Google AI Impact Challenge is a partnership between two different organizations, WattTime and The Carbon Tracker Initiative. We're a collection of data scientists, and project managers, and we think about the best ways for organizing our data and how best to engage new users.

What have you learned at Tech Sprint Week?

We’ve covered a lot! We went through a lot of user experience design and research, thinking about how users will be interacting with our product as we design it. We’ve also learned a lot about machine learning and feature engineering. The mentors reminded us to make sure we train our model on the type of data that it would actually have, which sounds intuitive but it's actually hard to do. It might be easy to give your model a “leg up” with training data that it shouldn't have, but then you would see that it's not operating as you expected. 

Now that Tech Sprint Week is complete, what are your next steps? 

We need to catch up with everybody else on the team and share all of the great information and resources that we've received from this week. I’ve also been exposed to a lot of new tools like TensorFlow, an open source library that makes it easy to create machine learning models. So I want to get familiar with that tool and actually integrate it into our workflow. We're also doing a lot of hiring, so we’ll continue to build our team. 

What kinds of people have you met through this program? 

All the mentors have been helpful. Everyone has this attitude of “Hey, how can we help?” Our AI Coach, Ang Li, has been extremely useful and really responsive. I'll contact him at random times of the day and get a response within a few minutes. 


Google for Chile: Supporting development through tech

Over the last decade, Chile has become known as one of the most connected countries in Latin America, and its population has been an early adopter of new technologies. But the country still has important challenges and opportunities to connect and bring all Chileans closer to technology that can make both their work and home lives easier.

Today we hosted our first Google for Chile, with a group of more than 300 people in Santiago. There, we discussed our ongoing commitment to the digital growth of Chile and Latin America, improving connectivity and creating a safer public cloud. 

Connecting Chile's entrepreneurial force

In Chile and around the world, small and medium businesses increasingly need to be online in order to grow. Google My Business has become one of the best allies for entrepreneurs who want to see their businesses "on the map" and for their customers to find them. The number of verified companies on the platform in Chile has grown by 76% over the past year.

More efficient cities, in the cloud

In Chile, almost 50 percent of drivers use Waze to drive around all types of streets. That means users can serve as a kind of “sensor” in addition to stationary ones like radar and cameras, and cities can learn a lot from their drivers. Now, all the information from the Waze for Cities program will be stored for free for its members on Google Cloud, making it even easier for cities to see movement patterns and measure the effects of interventions. Currently, more than 190 partners across Latin America have joined the program.  

Partners like the Subsecretaría de Transportes de Chile have been using Waze data to improve traffic. They monitor more than 400 road segments to determine the periods with the most traffic. This information is used to program traffic lights, and whenever patterns change (like when traffic piles up or there’s an accident on the road), they can adapt the lights accordingly. 

Keeping Chilean children, teachers and parents safe online

In 2018, we launched Be Internet Awesome, which teaches children to be safe explorers of the online world. In Chile, we have been working with the Education Ministry so teachers and administrators can use our program’s tools. In the coming weeks, teachers using Be Internet Awesome will be able to find a new module—in Spanish—to teach students to think critically about the information they consume online, avoiding misinformation. 

Privacy for all 

New privacy tools are now officially available in Chile. People can now use Android phones as security keys, adding an extra layer of protection to their information. They can also check how data is being used in Maps, Search and the Assistant, by accessing the apps menu and choosing the option “Your data in …” There, you can review and delete your location activity in Maps or your search activity in Search. Soon, the same feature will be accessible on YouTube.

Auto-delete controls for Web and Apps Activity are also now available globally, allowing people to easily manage the amount of time their data is saved. Choose a limit—3 or 18 months—and anything older than that will be automatically deleted on an ongoing basis.

The cloud in Quilicura

The first and only Google data center in Latin America is located in Chile, in the city of Quilicura. Announced in 2012, the data center allows us to provide support to and guarantee the operation of all of our products, not just for Chile but for all of Latin America. 

In September 2018, we announced the expansion of our data center, with an additional investment of US$140 million that will triple the size of the initial structure. And last April we announced the arrival of Curie on the coasts of the Valparaiso Region: Curie is the first submarine fiber optic cable to reach Chile in about 20 years.

How AI is transforming industries in Chile 

At Google, we use artificial intelligence to make our products more useful, from email that is spam-free and easier to write to a digital assistant that understands you when you ask it questions.

Much of the progress made with AI is based on our open source machine learning platform, TensorFlow. In Chile, machine learning is opening up new opportunities in several industries like food, construction and astronomy. Local technology company Odd Industries found potential in using AI with camera footage in the construction sector, letting data reveal what humans can’t see. Artificial intelligence processes images from construction sites and converts them into concrete data, allowing companies to build responsibly and intelligently. 

Working together with industry associations, academic institutions, government officials and our users, I’m excited to find new ways to use technology to help everyone succeed.