The leading autonomous driving company Waymo has teamed up with the AI specialist DeepMind to develop effective and efficient training methods to train its autonomous software.
Both Waymo and DeepMind are Alphabet companies. The partnership is aimed to leverage DeepMind’s Population-based training (PBT) method for Waymo’s self-driving cars.
Waymo self-driving cars use multiple neural nets and other methods to detect the pedestrians. Use of the PBT models can improve the ability of neural nets to detect pedestrians. The companies said that the goal of experimenting PBT is to train a single neural net so that it can maintain a recall over 99%. It could also decrease the false positives.
DeepMind detailed in a blog post that “our PBT models were able to achieve higher precision by reducing false positives by 24% compared to its hand-tuned equivalent, while maintaining a high recall rate.”
PBT can train and optimise the system. The software has neural nets that try an action and measure it to check if the actions are improving or not. This does not add any computational overhead and can be done as quickly as the traditional techniques do. Also, this new method of training is easy to merge into the existing machine learning (ML) pipelines.
“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,” added DeepMind team.
Previously, Waymo had a lot of neural nets that were working independently on the same task, all with a varied degree of deviation, also known as learning rate, each time they attempted a task.
Having a lower learning rate means a balanced and stable progress and a higher learning rate means a great variety when it comes to the quality of the outcome. The training to fill this gap takes a lot of effort as the engineers need to research manually to get rid of the low performing tasks.
PBT essentially is a self-driving technology that replaces the bad approaches of training with a better one, similar to the evolutionary theory.
“PBT uses half the computational resources used by random parallel search to efficiently discover better hyperparameter schedules. By incorporating PBT directly into Waymo’s technical infrastructure, researchers from across the company can apply this method with the click of a button, and spend less time tuning their learning rates,” said the researchers.