Updates from Coral

A better compiler, C++ API, updated OS, and board management tool

ic_share_24px Created with Sketch.
Close Created with Sketch.
Your Name, Developer Advocate, Google Research
April 26, 2019

“We saw that the only difference between cows that produce 30 liters of milk a day and those that produce 10 liters was the animal’s health. Could technology make cows healthier, and in doing so, help farmers grow their businesses?” Yasir recalls.

This question led him to start Connecterra, a company that uses our open-source machine learning framework, TensorFlow, to power what many call a “Fitbit for cows.” Using machine learning, Connecterra is able to diagnose problems early and provide recommendations to farmers on how to keep their cows healthy. Yasir’s story is an inspiring example of how machine learning can help tackle all kinds of problems. It also reminds us of the origins of TensorFlow itself—a product designed to help people everywhere make the most of AI.

The days of DistBelief

In 2011, when we first started working on machine learning in earnest at Google, the AI landscape was radically different. Computers had just become powerful enough to process what we then often called “big data,” and there were only early signs of progress in using deep learning for speech recognition. But with momentum around this modern machine learning building in academia, a handful of researchers and engineers at Google came together to form the Google Brain team. The group planned to focus on deep learning, a subset of machine learning that relies on using neural networks. Together, they developed our first internal ML system, which we called DistBelief. DistBelief allowed us to train complex models on large datasets and easily deploy those trained models into Google products. It quickly delivered results in Search and Google Photos, paving the way for the usage of machine learning company-wide.

Machine learning in Search lets people search in more natural, intuitive ways, using the language that they use everyday.

AI powers features in GMail like Nudging, Smart Reply and high-priority notifications, so people can spend more time on work that matters.

Google Translate uses a deep learning system called Neural Machine Translation to translate full sentences rather than translating queries word-for-word.

A fork in the roadmap

But by 2014, the original systems on which DistBelief was built were starting to change. Accelerators were replacing CPUs, and machine learning models were starting to become more complex, like sequence models for translation. Teams were finding that our internal systems couldn’t keep up with the ideas they had. And so we had a decision to make: we could either update DistBelief, knowing its structural limitations, or we could rebuild the system from the ground up. Then Brain team lead Jeff Dean had experience building big software systems at Google. He became a strong advocate for a fresh start to creating a machine learning platform — forming the basis of the system we now call TensorFlow.

There’s up to a 3% miss rate, even in the best hands... If you're going to put a patient through this semi-invasive procedure, you have to be as effective as you can,
Dr. Mankanwal Sachdev
Gastroenterologist at the Arizona Centers for Digestive Health

The thinking was that if we could share source code, it could accelerate progress in machine learning for everyone.
Rajat Monga
Director, TensorFlow

Machine learning for everyone

By design, TensorFlow is able to tackle a much wider range of machine learning problems than its predecessor. It allows users to design deep neural networks and run them on a single smartphone or across thousands of computers in data centers. As Jeff explains, “We wanted to keep the scalable attributes and production-readiness of our first system, but make it a much more flexible platform.” Building from scratch allowed us to take advantage of some of the exciting ideas we had seen externally and optimize for the newest hardware. Just a few months into the project, the team decided to make another ambitious choice: once TensorFlow was built and ready for production, they agreed they would open source it for anyone to use. “The thinking was that if we could share source code, it could accelerate progress in machine learning for everyone,” explained TensorFlow Director Rajat Monga.

Over the months that followed, teams across the company stress-tested and polished the framework until it was ready for release. TensorFlow quickly rose in popularity as a machine learning system at Google, powering ML implementations in products like Search, Gmail, Translate and more.

With over 50 internal teams using TensorFlow, we saw first-hand what it could do for our own products, but knew that these use cases were just the beginning. Since we open sourced the project in 2015, it’s become the most popular machine learning library on GitHub, with over 13 million downloads and over 1,300 contributors outside of Google. And each of those users comes with a story to tell. We’ve seen amazing work from researchers, developers, startups and enterprises across industries using TensorFlow to find creative solutions to a wide range of challenges.

Machine learning in Search lets people search in more natural, intuitive ways, using the language that they use everyday.