Waymo, a self-driving and robotaxi company, and DeepMind, an AI company, have teamed up to better train self-driving software in a way that mimics Darwinian evolution, according to a report.
Both companies are owned by Alphabet, which is also Google’s parent company. The companies use a method called population-based training (PBT), which lessens false positives when the software performs actions like placing boxes over the moving objects it sees in its sensors. The new training method also uses 50 percent less time and resources than previous methods.
PBT aims to help systems learn more efficiently. The neural nets in the software try an action and then measure it against a previous action to determine whether it was more “right” or “wrong,” the report noted.
In previous methods, Waymo would have a lot of neural nets working by themselves on the same task, and with a different degree of deviation, or learning rate, in the way they approached the task, whether it was for something like identifying foreign objects or stopping in a timely manner.
A lower learning rate means steady progress and a higher learning rate means more variety in the quality of the outcome. The comparative training takes a lot of work to get right, because engineers have to either use a “gut feeling” to determine what’s right or manually search results and get rid of the badly performing ones.
PBT essentially automates this process, killing off the bad training and replacing it with better approaches, similar to evolutionary theory.
DeepMind has been tweaking this method to get better results, and has also developed a method that would build sub-populations of neural nets and isolate them to see how they will develop on their own. Called the “island population” approach, this mimics the way animals who grow away from main populations develop different, but sometimes better, characteristics than their counterparts.