Partnering with Google has allowed DeepMind to bring the benefits of AI to billions of people worldwide. Through this collaboration, they have been able to apply breakthrough research to real-world problems at a global scale. One example of this is their partnership with Google Maps, which has had a significant impact on over one billion users.
Google Maps is relied upon by people for accurate traffic predictions and estimated times of arrival (ETAs). These features are essential for navigating through traffic jams, notifying others of delays, and ensuring timely attendance at important meetings. Businesses like rideshare companies also depend on these features to provide their services efficiently.
To improve the accuracy of real-time ETAs, DeepMind has worked with the Google Maps team using advanced machine learning techniques, including Graph Neural Networks. This partnership has resulted in up to a 50% improvement in ETA accuracy in cities like Berlin, Jakarta, São Paulo, Sydney, Tokyo, and Washington D.C.
Google Maps calculates ETAs by analyzing live traffic data, but it also incorporates historical traffic patterns to predict future traffic conditions. This process is complex due to factors such as varying rush-hour times, road conditions, accidents, and closures. DeepMind’s collaboration with Google Maps has helped minimize inaccuracies in ETAs, further enhancing the reliability of the service.
To achieve this improvement, DeepMind has used a machine learning architecture called Graph Neural Networks, which allows for spatiotemporal reasoning by modeling the connectivity structure of real-world road networks. They divided road networks into “Supersegments,” which are groups of adjacent road segments that share significant traffic volume. By training a single Graph Neural Network model using these Supersegments, they were able to deploy this solution at a global scale.
Graph Neural Networks enable modeling of network dynamics and information propagation, making them suitable for modeling traffic flow. By extending the learning bias of previous neural network architectures, Graph Neural Networks can handle complex connections and predict delays and traversal times accurately.
Adapting Graph Neural Networks for production-ready machine learning models required addressing the variability that can exist across multiple training runs. DeepMind developed an approach using reinforcement learning techniques adapted for a supervised setting. This helped overcome the instability caused by variable graph structures used during training.
Overall, DeepMind’s partnership with Google Maps has significantly improved the accuracy of ETAs, benefiting users around the world. This collaboration showcases the power of AI and its ability to address real-world problems on a global scale.