World-scale Deep Learning for Automated Driving
World-scale Deep Learning for Automated Driving

Abstract: 

Automated driving solutions have mostly resorted to supervised learning as their primary mode for training of ML models for robust perception. As a consequence, the autonomous vehicle industry has been focused on amassing large volumes of labeled data to have a competitive edge in the deep-learning era. At TRI, we realize the need to go beyond supervised learning for automated driving, especially in computer vision problems that are seeing great progress with strong supervision today. First, we will motivate an exciting scientific problem of self-supervised learning that can have huge implications in the research and development of autonomous robots at world-scale. Second, we will discuss some of the recent state-of-the-art techniques and results we have developed at TRI, specifically on the self-supervision of Pseudo-Lidar Networks. Finally, we will showcase some of the exciting large-scale training infrastructure we leverage in order to solve these scalable-but-challenging self-supervised techniques.

Bio: 

Sudeep is a Manager and Research Scientist in the Machine Learning Team at Toyota Research Institute. He received his PhD in Computer Science from MIT (CSAIL), where he focused on self-supervised perception and learning in SLAM-aware mobile robots. Prior to MIT, he was a software developer working on real-time computer-vision and motion-capture related technologies.