NVIDIA adopts federated learning to train autonomous vehicles across borders

Federated learning has proven to be a game-changer in autonomous vehicle (AV) development, especially in scenarios spanning multiple countries. This innovative approach allows the use of a variety of data sources and conditions that are critical to improving AV technology. According to the NVIDIA technology blog, federated learning allows AVs to jointly train algorithms with locally collected data, maintain data decentralization, and enhance privacy and security.

Strengthen privacy and compliance

Unlike traditional machine learning methods that require centralized data storage, federated learning ensures that sensitive data remains within the country of origin. This approach not only enhances privacy protection but also complies with various international data protection regulations, such as the European Union’s GDPR and China’s PIPL. Federated learning helps AVs comply with these regulations by minimizing data movement while benefiting from collective learning processes.

NVIDIA Federated Learning Platform

NVIDIA developed its AV federated learning platform using the open source framework NVIDIA FLARE. The platform allows you to train global models by integrating data from multiple countries, solving regulatory and logistical challenges associated with traditional centralized data processing.

The deployment setup consists of two federated learning clients and a central server, with the FL server hosted on AWS in Japan. The system integrates with existing AV machine learning infrastructure to facilitate seamless data processing and model training.

Motivation and Use Cases

The NVIDIA AV team operates on a global scale and collects data from various regions to improve AV capabilities. The need to handle data from multiple countries stems from the need to address rare use cases that may not exist in all countries. The platform supports tasks such as object detection and symbol recognition, enabling the development of integrated global models that meet or exceed the performance of individual country-specific models.

Challenges and Solutions

Implementing a global AI model requires several challenges, including IT setup, network bandwidth, and disruption. NVIDIA solved these problems by hosting FL servers on AWS and optimizing the model transfer process. The team also implemented solutions to recover from network outages, ensuring uninterrupted training sessions.

Project status and future prospects

Since the platform was deployed, the number of data scientists has grown from 2 to 30. NVIDIA has used this platform to successfully train and launch numerous AV models, demonstrating excellent performance on tasks such as road sign recognition.

This federated learning approach not only improves model training without data movement, but also ensures compliance and cost-effectiveness. NVIDIA’s strategy of developing this platform can be applied to other industries such as healthcare and finance, further expanding the scope of federated learning applications.

Image source: Shutterstock