The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Advancing Autonomous Vehicle Development Using Distributed Deep Learning with Adrien Gaidon - TWiML Talk #269
May 28, 2019
Adrien Gaidon, Machine Learning Lead at Toyota Research Institute, shares his journey into distributed deep learning for autonomous vehicles. He dives into the evolution of TRI's platform and the pivotal role of data simulation in this field. Gaidon discusses the complexities of building infrastructure for large datasets and GPU management, as well as the integration of PyTorch and Horovod. He highlights advancements in model compression and multitask learning, emphasizing their importance for efficient self-driving technology.
AI Snips
Chapters
Transcript
Episode notes
From On-Premise to Cloud
- Adrien Gaidon initially planned to buy many GPUs for on-premise deep learning at Toyota Research Institute (TRI).
- However, scaling proved slow, leading to exploration of cloud-based solutions with AWS.
Unique Needs of Autonomous Driving
- Autonomous driving necessitates small, high-resolution networks for real-time performance and long-range prediction.
- Standard deep learning tools often don't scale efficiently for these specific requirements.
Building from Scratch
- TRI started with single-node setups, leveraging large RAM and optimizing for speed over cost.
- They transitioned to a BeeGFS-based distributed file system to handle large datasets and improve data access.