Deep Driving: End-to-End Learning Strategies for Self-Driving Cars

ABSTRACT

The canonical way of developing self-driving cars involves a modular approach, where components such as perception, prediction, planning, and control are developed separately. An end-to-end learning strategy for self-driving cars aims to map sensory data directly to driving commands with just one functional component, such as a neural network. In this talk, I will present an overview of deep learning, particularly convolutional neural networks, for a physics audience, a few versions of such neural networks we developed for end-to-end driving, and strategies of training and testing them. 

BIOGRAPHY

Xinwei "Sam" Gong has Bachelor's degrees in mathematics and physics from Virginia Tech, and a PhD in physics from Duke University. His PhD research involved dynamics of Boolean networks and statistical complexity of spatially extended systems. Sam acquired core machine learning knowledge in his graduate studies, and has been working on machine learning problems in Silicon Valley after graduation. Sam is currently a Senior Staff Machine Learning Engineer at a research lab of Volkswagen Group, working on self-driving technologies.