Cutting edge technology

Introduction: On January 16th, US local time, Waymo, the global autopilot leader, released a short video of the 10-year anniversary celebration and published an article on “Auto ML (Auto Machine Learning)” on the official blog. Anatomy of how Auto ML and Google AI brains help Waymo develop autonomous driving technology.

On January 16th, US local time, Waymo, the global autopilot leader, released a short video of the 10-year anniversary celebration and published an article on "Auto ML (Auto Machine Learning)" on the official blog. An in-depth analysis of how Auto ML and Google AI brains help Waymo develop autopilot technology.

wrote on Waymo's official tweet: This week, ten years ago, the “Project Driver” was formally established with the mission of improving road safety and making transportation more convenient. From this "moon landing" project to the Google Autopilot project, now is Waymo, working together for the next decade and beyond!

Here's an article about Auto ML, where machine learning plays a key role in almost every part of an automated driving system. It helps our cars see the environment around them, understand the world, predict the behavior of others, and decide on the best course for their next step.

In the case of perception, Waymo's system uses a combination of neural networks that enables Waymo's vehicles to interpret sensor data, identify objects, and track them over time, thus providing an in-depth view of the world around them. Understanding.

Creating these neural networks is often a time-consuming task: optimizing the neural network architecture to achieve the quality and speed required to automate a car's operation is a complex fine-tuning process that Waymo engineers may need for months Time to complete a new task.

Now, in collaboration with researchers from the Google AI brain, Waymo is putting cutting-edge research into practice to automatically generate neural networks. More importantly, these state-of-the-art neural networks are of higher quality and faster than those manually adjusted by engineers.

In order to apply Waymo's autonomous driving technology to different cities and environments, it is necessary to quickly optimize the Waymo model for different scenarios. Auto ML enables Waymo to do this, providing a large number of ML solutions efficiently and continuously.

01 Migration Learning: Using Existing Automation Architectures

The collaboration between Waymo and Google AI brain begins with a simple question: Can Auto ML generate high quality, low latency for cars? Neural Networks?

The standard for quality measurement is the accuracy of the answers generated by the neural network, and the rate at which the delay measure network provides answers, also known as reasoning time. Since driving is an activity, it requires the vehicle to use real-time answers, and considering the security of the system, the neural network needs to operate with low latency. Most networks run directly on Waymo's vehicles, resulting in less than 10 milliseconds, which is faster than many of the networks in a data center deployed on thousands of servers.

In the original Auto ML (Learning Transferable Architectures for Scalable Image Recognition PDF), Google AI employees can automatically explore more than 12,000 architectures to solve the CIFAR-10 classic image recognition task: identify a The small image represents one of ten categories, such as buying a car, an airplane, a dog, and so on.


PDF, at the end of the article), they found a building block of the family's neural network, called the NAS unit, which may be Automatically build tasks like manual network CIFAR-10 and similar. Through this collaboration, Waymo's researchers decided to use these units to automatically build new models for autonomous driving tasks, thereby transferring knowledge on CIFAR-10 to the automotive field. The first experiment was a semantic segmentation task: Identifying lidar points Every point in the cloud, such as cars, pedestrians, trees, etc.

Figure 1: A NASAn example of a unit that processes the first two layers of input in a neural network

To this end, Waymo researchers have built an automated search algorithm that explores hundreds of different types in a convolutional network architecture (CNN). NAS unit combination, which is Waymo's lidar segmentation task training and evaluation model. When Waymo engineers manually tweaked these networks, they could only explore a limited number of architectures, but with this approach, hundreds of architectures could be explored automatically.

Compared to previous artificial fine tuning optimization neural networks, Auto ML is improved in two ways:

some delays with similar quality are significantly reduced;

others Has a higher quality and similar delay.

After initial success, Waymo applied the same search algorithm to two other tasks related to traffic lane detection and location. Transfer learning techniques are also applicable to these tasks, and finally three new trainings can be deployed on the car. And improved neural networks.

Waymo self-driving car (Prius) ten years ago

This article was written by the author of the cutting-edge technology. The idea is only the author and does not represent the OFweek position. If you have any infringement or other problems, please contact us.

Hot topic