Much work has been done in exploring deep learning for the perception task of autonomous driving, specifically through the use of convolutional neural networks (CNNs) for object recognition and lane following. CNNs have also been explored in NVIDIA’s paper on End to End Learning for Self-driving Cars, where raw pixels from a front-facing camera were mapped to steering commands. How much potential is there in applying other deep learning architectures like Recurrent Neural Networks and Deep Reinforcement Learning to autonomous driving, what has already been done, and what are the challenges?
Jia Qing is the Founder, Executive Lead and Deep Learning Engineer of OpenSourceSDC, where they are opening up testing vehicles for anyone to test code for autonomous driving. He believes that future breakthroughs in the technology lie in bringing in techniques from other domains, and that can be achieved by bringing accessibility of testing vehicles to people in tangentially-related domains like computational neuroscience, industry automation, data science, signal processing, all in an open-source fashion. Jia Qing has been looking into deep learning methods in natural language processing and computer vision, and is more recently excited about the potential of deep reinforcement learning and generative adversarial networks.