Waymo, a branch of Google’s Alphabet which is currently developing self-driving cars and taxis, has announced that it’s been working with DeepMind on a new technique for training its autonomous vehicles.
Waymo and AI specialist DeepMind, which is also an Alphabet company, created a “Population Based Training” (PBT) method to train Waymo’s self-driving cars in a more efficient way.
The PBT method “enabled dramatic improvements in how well -- and how quickly -- vehicles were able to sense pedestrians, bicyclists, and motorcycles,” Waymo stated. The approach decreased false positives by 24 percent compared to the AI algorithm that involves having researchers and engineers handpick AI models that need improvement through trial and error.
Evolution-based training method
The method has neural networks (which Waymo’s self-driving cars use to execute many driving tasks) “compete with each for ‘survival’ in an evolutionary fashion,” explained Yu-hsin Chen, Waymo’s lead software engineer in a blog post.
“If a member of the population is underperforming, it’s replaced with the ‘progeny’ of a better performing member,” Chen said. “PBT doesn’t require us to restart training from scratch, because each progeny inherits the full state of its parent network, and hyperparameters are updated actively throughout training, not at the end of training.”
“Compared to random search, PBT spends more of its resources training with good hyperparameter values,” Chen said.
In an interview with the Financial Times, Chen said his team saw improvement in tasks like "detecting pedestrians, cyclists and motorcyclists, highway lanes, vegetation, the road, and it also improved our data labelling process."