Explaining How a Deep Neural Network Trained with End-to-End Learning Steers a Car

By Mariusz Bojarski, Philip Yeres, Anna Choromanska, Krzysztof Marcin Choromanski, Bernhard Firner, L. Jackel, Urs Muller. 2017

437 citations recorded by the source corpus.

As part of a complete software stack for autonomous driving, NVIDIA has created a neural-network-based system, known as PilotNet, which outputs steering angles given images of the road ahead. PilotNet is trained using road images paired with the steering angles generated by a human driving a data-collection car. It derives the necessary domain knowledge by observing human drivers. This eliminates the need for human engineers to anticipate what is important in an image and foresee all the necessary rules for safe driving. Road tests demonstrated that PilotNet can successfully perform lane keeping in a wide variety of driving conditions, regardless of whether lane markings are present or not. The goal of the work described here is to explain what PilotNet learns and how it makes its decisions. To this end we developed a method for determining which elements in the road image most influence PilotNet's steering decision. Results show that PilotNet indeed learns to recognize relevant objects on the road. In addition to learning the obvious features such as lane markings, edges of roads, and other cars, PilotNet learns more subtle features that would be hard to anticipate and program by engineers, for example, bushes lining the edge of the road and atypical vehicle classes.

Subfields: Artificial Intelligence, Automotive Engineering, Computer Vision and Pattern Recognition, Safety, Robustness & Interpretability, Robotics & Control

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Automotive Engineering

In Computer Vision and Pattern Recognition

In Safety, Robustness & Interpretability

In Robotics & Control

Citations in the corpus

When Does Machine Learning FAIL? Generalized Transferability for Evasion and Poisoning Attacks (2018, background): “the justice system [31], cybersecurity [1], or self-driving cars [6].”
Self-Driving Car Steering Angle Prediction Based on Image Recognition (2019, methodology): “A similar approach was used in a recent NVIDIA paper [2].”
Peeking Inside the Black-Box: A Survey on Explainable Artificial Intelligence (XAI) (2018, neutral): “Available: https://arxiv. org/abs/1703.04730 [24] M. Bojarski et al. (2017).”
Explanation Methods in Deep Learning: Users, Values, Concerns and Challenges (2018, background): “, 2017), autonomous driving (Bojarski et al., 2017) and legislation (Lockett et al.”
Methods for interpreting and understanding deep neural networks (2017, background): “Interpretability is especially important in applications such as medicine or self-driving cars, where the reliance of the model on the correct features must be guaranteed [15, 14].”
Deep Reinforcement Learning for Autonomous Driving: A Survey (2020, background): “An architecture for learning a convolutional neural network, end to end, in self-driving cars domain was proposed in [114], [115].”
Agile Autonomous Driving using End-to-End Deep Imitation Learning (2017, background): “Note this control frequency is significantly higher than [2] (10 Hz), [24] (12 Hz), and [19] (15 Hz).”
Concrete Problems for Autonomous Vehicle Safety: Advantages of Bayesian Deep Learning (2017, methodology): “When applied to autonomous vehicle models, research shows that driving decisions depend most on the boundaries of other cars and lane markings immediately ahead [Bojarski et al., 2017].”
Agile Off-Road Autonomous Driving Using End-to-End Deep Imitation Learning (2017, methodology): “Therefore, we conjecture state-of-the-art CNNs (e.g. [7]) cannot be directly used in this task.”
Unsupervised Real-to-Virtual Domain Unification for End-to-End Highway Driving (2018, methodology): “[6] visualizes the activation of the neural network and shows that the model not only learns driving-critical information such as lane markings, but also unexpected features such as atypical vehicle classes.”
Simba: Scaling Deep-Learning Inference with Multi-Chip-Module-Based Architecture (2019, methodology): “We use DriveNet to demonstrate the weak scaling trend of Simba.”
An End-to-End Deep Neural Network for Autonomous Driving Designed for Embedded Automotive Platforms (2019, methodology): “PilotNet, also called NVIDIA CNN that is designed by researchers from NVIDIA Corporation, is the deep neural network for end-to-end autonomous driving, created with the motivation to improve the DARPA Autonomous Vehicle (DAVE) system for autonomous driving [1,2].”
End-to-End Learning of Driving Models with Surround-View Cameras and Route Planners (2018, background): “Methods have also been developed to explain how the end-to-end networks work for the driving task [30] and to predict when they fail [17].”
Early Failure Detection of Deep End-to-End Control Policy by Reinforcement Learning (2019, background): “The NVIDIA PilotNet project [4] does consider vision-based vehicle control with end-to-end learning and learns to recognize diverse obstacles and objects, but is not a probabilistic method and does not seem to explicitly handle truly novel test inputs.”
A survey of deep learning techniques for autonomous driving (2019, background): “Through training, PilotNet learns how the steering commands are computed by a human driver (Bojarski et al., 2017).”
Artificial intelligence applications in the development of autonomous vehicles: a survey (2020, methodology): “[20] Camera (left, center, right) Y Y Y Y Y Y SimulationOn road test (10-mile) CNN”
Neural Networks and Deep Learning (2018, neutral): “The description of self-driving cars is based on [46, 47].”
ChauffeurNet: Learning to Drive by Imitating the Best and Synthesizing the Worst (2018, background): “Researchers at NVIDIA (Bojarski et al. (2016, 2017)) showed how to train an end-to-end deep convolutional neural network that steers a car by consuming camera input.”
Interpretability of deep learning models: A survey of results (2017, background): “Computational models that impart reasoning behind their decisions, often use the terms “interpretability” and “explainability” synonymously [16], [47], [48].”