Alphabet’s multi-billion dollar subsidiary Waymo plays a leading role in the modern autonomous driving industry, and the company’s cutting-edge self-driving system has allowed it to maintain this high valuation on Wall Street.
Machine learning (ML) drives every part of the Waymo self-driving system. Is the core method that enables self-driving vehicles to visualize their surroundings, sense the world, predict behaviors, and calculate actions and movement etc.
To make ML-based solutions available for a wider variety of deployment scenarios, Waymo’s autonomous driving team has collaborated with Google AI Brain Team researchers on a system that automates the creation of high quality and low latency neural networks on existing AutoML architectures.
Conduct Transferable Learning Models on AutoML Architectures
Building on its previous research with Neural Architecture Search (NAS), the team used NAS cells to automatically build new machine learning models for tasks specific to self-driving, and transferred CIFAR-10 model learning results directly to the system. The team then set up an automatic search algorithm to enable different NAS cell combinations with convolutional neural networks (CNN) for training and evaluating of suitable models for LiDAR point identification segmentation tasks. With its success in LiDAR tasks, the team leveraged the transfer learning techniques on LiDAR detection and localization of traffic lanes tasks, which also produced favourable results.
New End-to-End Searching Architectures
Waymo and Google Brian researchers took the system a step further, looking for completely new architectures that could improve results while keeping latency low. Conducting an end-to-end search using traditional methods is very time-consuming, so researchers came up with a proxy task which could solve scaled-down LiDAR segmentation tasks in a fraction of the time. By exploring more than 10,000 different architecture with the proxy task, the team was able to build neural nets with 20-30 percent lower latency and 8-10 percent lower error rates at the same latency as previous architectures.
Both transfer learning and proxy end-to-end search methods provided nets that were better than hand-crafted ones for the LiDAR segmentation tasks for autonomous driving. The team intends to continue its collaboration, confident the research on automatically building ML architectures can also be used to optimize neural nets for additional task types.
Author: Robert Tian | Editor: Michael Sarazen