SUBARU Corporation: Turns to Google Cloud AI and machine learning to accelerate the AI development for EyeSight
About SUBARU Corporation
SUBARU is widely recognized as an automobile brand name and the organization manufactures and sells vehicles, aircraft, and space-related equipment. As of March 2021, the business had 36,000+ employees.
Tell us your challenge. We're here to help.
Contact usWith Google Cloud, SUBARU Corporation is developing AI, machine learning, and deep learning to improve its advanced driver-assistance system in support of a program to eliminate fatal traffic accidents caused by its own cars by 2030.
Google Cloud results
- Reduces pre-processing time to add annotation data to image data for learning and generating TensorFlow records from more than a day to 30 minutes
- Prevents the leakage of sensitive information
- Positions the business to implement MLOps for effective machine learning deployment and operation
Reduces data processing time from several days to 30 minutes
SUBARU Corporation aims to achieve the target of zero fatal traffic accidents1 by 2030 and opened an AI development base, SUBARU Lab at Shibuya, Tokyo in December 2020, to accelerate research and development to further improve the safety of its EyeSight driver support system.
SUBARU Corporation first introduced stereo camera-based driving support technology Active Driving Assist (ADA) in the Legacy Lancaster in 1999 and released a more advanced version, EyeSight, in 2008. EyeSight became a key reason for customers to choose SUBARU when the organization released the second generation model in 2010.
However, while the technology is widely recognized as a leading advanced driver assistance system, deep learning—a technology and method that gained momentum in the latter half of the decade—presented an opportunity to provide an even better experience to drivers.
According to Toshimi Okubo, SUBARU Corporation Senior Engineer of AI R&D Section, ADAS Development Department, Engineering Division, "We have been researching deep learning since around 2015, and recently we started focusing on developing AI for EyeSight.”
However, the workstations used for development were completely inadequate for full-scale research into deep learning. Toshimi considered replacing these workstations, but the time and effort required to introduce new workstations and the fact these would not fully resolve the problem of development speeds not meeting requirements ruled this option out. SUBARU instead opted for the cloud.
“I chose Google Cloud from many platforms because, at the time of selection, it had multiple managed services such as Vertex AI, the managed notebooks option, and Vertex AI Training that were useful for AI development. It was also fascinating to have high-performance hardware that could handle large-scale machine learning operations.”
—Toshimi Okubo, SUBARU Corporation Senior Engineer of AI R&D Section, ADAS Development Department, Engineering DivisionMultiple managed AI services prompt Google Cloud selection
“The idea was to solve this problem by utilizing a cloud platform that could develop AI models and perform ML training at high speed,” says Toshimi “I chose Google Cloud from many platforms because it had multiple managed services such as Vertex AI, the managed notebooks option, and Vertex AI Training that were useful for AI development. It was also fascinating to have high-performance hardware that could handle large-scale machine learning operations.” As an example, Toshimi shared that Vertex AI notebooks reduce the time needed to setup the JupyterLab environment and are able to seamlessly integrate with Google Cloud services such as Cloud Storage and GPU instances.
SUBARU points to the combination of NVIDIA A100 GPUs with Compute Engine, and the provision of 16 of these GPUs in a single virtual machine, as important to its machine learning development because deep learning calculations for recognizing images need multiple numbers of the latest GPUs, whose performance is several times better than past models.
SUBARU Lab now uses Google Cloud to analyze a considerable number of images taken by EyeSight stereo cameras, while the multitasking neural network SUBARU ASURA Net performs tasks such as the detection of objects including cars and people and semantic segmentation that associates labels and categories with pixels in images and recognizes safe, drivable areas.
SUBARU initially used Google Cloud as infrastructure as a service due to concerns over security and responsiveness due to the distance between the Google region and the office location. Utilizing Identity-Aware Proxy (IAP) to control access to Google Cloud for security showed that response time was not the issue, because Vertex AI services respond very quickly. “Since I started actively using managed services such as Vertex AI, my team has worked entirely in the cloud,” says Toshimi. Each person can freely experiment, train, and create tasks through Vertex AI notebooks and the business does not have to concern itself with the management of hardware resources.
The organization continues to run on-premises environments with enhanced functions, meaning its overall operation runs in a hybrid environment. Petabyte-class image and video data remains on-premises while SUBARU Lab determines whether the returns from moving these to the cloud are commensurate with the migration costs.
“I was most happy with the introduction of Google Cloud, which almost eliminated the budgeting and internal coordination associated with strengthening the development environment. Thanks to that, I was able to concentrate on development.”
—Toshimi Okubo, SUBARU Corporation Senior Engineer of AI R&D Section, ADAS Development Department, Engineering DivisionCloud IAP delivers improved security
SUBARU Lab uses Cloud Source Repositories to manage code across the cloud and on-premises environments and has built a mechanism that automatically synchronizes code when the SUBARU ASURA Net model is updated or functions are added on-premises. In addition, the organization uses IAP to deliver improved security. “In this environment, to prevent information leakage, we are designing and building on the premise that external IP addresses are not able to access Compute Engine instances—with access provided instead through IAP,” says Takashi Kanai, SUBARU Corporation Manager of ADAS Development Department, Engineering Division. Furthermore, the organization uses Dataflow to process data at scale for this initiative.
Below is an image of the SUBARU Google Cloud architecture:
“It would have taken several months to prepare the equipment because of the delays. With Google Cloud, we can prepare a development environment that same day, so the speed is considerably faster.”
—Takashi Kanai, SUBARU Corporation Manager of ADAS Development Department, Engineering DivisionAutoscaling helps cut pre-processing times
“Pre-processing to add annotation data to image data for learning and generating TensorFlow records increases day by day, so it takes more than a whole day even when undertaken in parallel with conventional methods,” says Toshimi. “Therefore, I am trying to process these with Dataflow using Apache Beam—as a result, when data flows, our infrastructure automatically scales to several hundred CPUs at once and the process takes about 30 minutes.”
SUBARU Lab continues to work towards the company’s “zero fatal traffic accidents1 by 2030” target and Google Cloud is helping achieve this target. “I was most happy with the introduction of Google Cloud, which almost eliminated the budgeting and internal coordination associated with strengthening the development environment,” says Toshimi. “Thanks to that, I was able to concentrate on development.”
The organization hires engineers on a monthly basis, but particularly during the pandemic, would have found it difficult to prepare a development environment for each recruit. “It would have taken several months to prepare the equipment because of the delays,” says Takashi. “With Google Cloud, we can prepare a development environment in one day, so the speed is considerably faster.”
In the future, SUBARU Lab plans to promote establishment of an MLOps system and infrastructure to further streamline AI development. “For that purpose, we are considering and trying various Google Cloud services that implement MLOps,” says Takashi.
1. Reducing to zero the number of fatal accidents occurring while a driver or passenger in a SUBARU and the number of fatalities among pedestrians, cyclists, and the like arising from collisions with a SUBARU vehicle.
Tell us your challenge. We're here to help.
Contact us
About SUBARU Corporation
SUBARU is widely recognized as an automobile brand name and the organization manufactures and sells vehicles, aircraft, and space-related equipment. As of March 2021, the business had 36,000+ employees.