Deepmind’s ML Efforts make Google Maps get Better


Google maps get 20million pieces of information every day due to the google users in every second. The problem of traffic can crash the algorithms predicting the ETA and also there are chances of new roads and buildings built all the time. Even though Google maps gets its ETA all the time correct but still there is a need for improvement.

Researchers at Deepmind have partnered with the google map to improve its accuracy of real-time ETAs up to 50% in places like Berlin, Sydney, Sao Paulo, Tokyo, and Washington DC. They are performing this by using advanced machine learning techniques which include Graph Neural Networks.

How deep mind worked out a plan

Road networks were divided into super segment consisting of multiple adjacent segments which share significant traffic volume and google map calculates its ETAs by analyzing live traffic for road segments. But it doesn’t consider how much traffic a driver can expect in 10 minutes during the drive. To solve this deep mind researchers used graph neural networks which is a type of machine architecture for spatiotemporal reasoning and this architecture incorporated relational learning biases to model. They are trained as a single, fully connected neural network model for all super segments. To incorporate this in scale millions of models need to be trained, hence the researcher decided to use graph neural networks.

In graph neural networks, super segments are road subgraphs and are sampled randomly in proportion to traffic density. Hence, a single model is used to train these sampled subgraphs and can deploy at a scale. Graph neural networks expand the learning bias imposed by the conventional neural networks and recurrent neural networks by using the concept of proximity, which can handle traffic on adjacent and intersecting roads. This ability to graph neural networks to generalize over combinatorial spaces helps the modeling technique to gain power.

Researchers also discovered that graph neural networks are sensitive to changes in the curriculum, hence to solve the problem of variability in graph structures a novel reinforcement learning technique was provided. Deepmind implemented the method of meta gradients to adjust the learning rate to the training and in this way, the system learned its learning rate schedules. Besides, the result showed that ETA inaccuracies have decreased significantly.

Google maps keep getting better

Due to the machine learning algorithm, google map services kept getting better. In most countries like India google has introduced its services like to track bus routes, even in the densely populated neighborhood the model was able to predict vehicle speed accurately. Models were also designed to acquire unique properties of specific streets, cities, etc.

While the ultimate aim of the machine learning model was to reduce errors in the travel estimates, the Deepmind researchers also found that using a linear combination of multiple loss functions can also greatly increase the ability of the model to generalize.


Please enter your comment!
Please enter your name here