The process of training self-driving vehicle AI is seldom efficient when you need to either use a massive quantity of computing power to train techniques in parallel or else have researchers spend ages manually weeding out dangerous methods. Waymo might need a better approach: use the identical principles that guide evolution. The corporate has partnered with DeepMind on a “Population-Based Training” technique for pedestrian detection that has the very best neural networks advance much like lifeforms do in natural selection, saving effort and time.
The approach regularly has the networks compete against each other, with weaker examples being changed by stronger “progeny” which can be copies of the better-performing networks with barely tweaked parameters (just as a baby is not a perfect clone of its parent). This automatically eliminates the poorer-performing networks while saving Waymo from having to retrain networks from scratch- they’ve already inherited technology from their parents.
There’s a risk that the method is targeted too much on short-term improvements. To fight this, Waymo developed “niches” where neural networks challenged each other in sub-groups to get strong outcomes while preserving range that could be better suited for real-world driving situations.
The outcomes have been promising when applied to pedestrian detection. The PBT strategy dropped false positives by 24 %, regardless that it took half as a lot of time. The experiment went so great that Waymo has even been utilizing PBT throughout other models. That, in turn, guarantees self-driving vehicles that can better cope with the complexities of driving and avoid collisions.