We work closely with Google and other experts to find ways for our advances to benefit society.
So far, our systems have shown how they can save energy, identify eye disease, accelerate science, and improve Google products used across the world.
Identifying eye disease faster
We partnered with Moorfields Eye Hospital to develop faster ways of identifying, and better ways of understanding, common eye diseases from routine scans.
Over 100 million people are affected by diabetic retinopathy or age-related macular degeneration. These conditions can cause permanent sight loss unless they’re treated quickly. The results, which were published in Nature Medicine, showed that our AI system could recommend patient referrals as accurately as world-leading expert doctors for over 50 sight-threatening eye diseases. More recently, we showed that our system can predict whether a patient will develop a more severe form of age-related macular degeneration months before it happens–paving the way for future research in sight-loss prevention.
Saving energy at Google scale
Google's data centres contain thousands of servers that power services including Google Search, Gmail, and YouTube.
It’s essential to keep these servers cool, but this takes a lot of electricity. Even minor improvements would significantly reduce energy usage and CO2 emissions. By building an AI system that manages data centre cooling more efficiently, we helped save around 30% of the energy needed. We also designed it with safety and reliability in mind, so the AI system works within strict constraints.
Improving Google products
Working with Google, we’ve applied our cutting-edge research to products and infrastructure used by billions of people.
Our voice synthesis technology WaveNet is in the hands of people who use Google Assistant and Google Cloud Platform around the world. And our systems have helped improve mobile phone battery use and screen brightness for millions of people using the Android Pie operating system.
Understanding protein folding
Proteins are complex molecules that are essential to life, and each has its own unique 3D shape.
Knowing how proteins fold to create different shapes could help scientists understand a protein’s role within the body. This discovery might help treat diseases believed to involve misfolded proteins such as Parkinson’s, Huntington’s and cystic fibrosis. Predicting the shape of proteins is a major unsolved challenge in science and we’ve already seen early signs that our AI systems could accelerate progress in this field.
These real-world projects build on our breakthroughs in fundamental AI research, published in journals like Nature, Science, and more.
AlphaZero: Shedding new light on chess, shogi, and Go
AlphaZero is a single system that learned to play three famously complex games, becoming the strongest player in history for each. Learning entirely from scratch, AlphaZero developed its own distinctive style that continues to inspire human grandmasters.
DQN: Human-level control of Atari games
One of the great challenges in AI is building flexible systems that can take on a wide range of tasks. Our Deep Q-Network (DQN) surpassed the overall performance of professional players in 49 different Atari games using only raw pixels and the score as inputs.
A neural network with dynamic memory
The differentiable neural computer (DNC) is a system that learns to use its external memory to answer questions about different kinds of complex structured data, such as artificially generated stories, family trees, or a map of the London Underground.
AlphaGo defeats Lee Sedol in the game of Go
While becoming the first computer program to defeat a professional human Go player, AlphaGo taught the world new knowledge about perhaps the most studied and contemplated game in history.
WaveNet: A generative model for raw audio
WaveNet generates realistic human-sounding speech that reduced the gap between computer and human performance by over 50%, when it was introduced.
GQN: Neural scene representation and rendering
The Generative Query Network (GQN) allows computers to learn about a generated scene purely from observation, much like how infants learn to understand the world.