Jul 30, 2019

Waymo and DeepMind use principles of human evolution to improve autonomous driving

The two companies in the Alphabet group have created an approach that promises better results and faster training in artificial intelligence systems


Population Based Training (PBT) is one of the methods that Waymo specializes in driving solutions and autonomous vehicles, and DeepMind, focused on artificial intelligence, are using to train algorithms that create virtual conductors. In the first tests, the PBT created 24% less false positives in a network that identifies pedestrians, cyclists and motorbikes through the various sensors of WAYMO. The training time and resources used were still 50% lower than with the methods currently used by Waymo.

PBT automates the process of selecting neural networks that have worse performances, replicating the principles of natural evolution listed by Charles Darwin in his evolutionary theory, explains TechCrunch.

Researchers from both companies have improved the natural selection process by creating rapid evaluation models, with 15-minute intervals, built robust validation criteria and realistic sample sets to ensure that neural networks were adapted to reality.

This approach, as it is faster and more effective, promises to contribute to the evolution of artificial intelligence learning systems in autonomous driving.