How quickly can a computer make sense of what it sees without losing accuracy? And to what extent can AI tasks on hardware be performed with limited computing resources? Aiming to answer these and other questions, car-safety software company Zenseact, founded by Volvo Cars, sought out CERN’s unique capabilities in real-time data analysis to investigate applications of machine-learning to autonomous driving.
In the future, self-driving cars are expected to considerably reduce the number of road-accident fatalities. To advance developments, in 2019 CERN and Zenseact began a three-year project to research machine-learning models that could enable self-driving cars to make better decisions faster. Carried out in an open-source software environment, the project’s focus was “computer vision” – an AI discipline dealing with how computers interpret the visual world and then automate actions based on that understanding.
“Deep learning has strongly reshaped computer vision in the last decade, and the accuracy of image-recognition applications is now at unprecedented levels. But the results of our research show that there’s still room for improvement when it comes to running the deep-learning algorithms faster and being more energy-efficient on resource-limited on-device hardware,” said Christoffer Petersson, research lead at Zenseact. “Simply put, machine-learning techniques might help drive faster decision-making in autonomous cars.”
The need to react fast and make quick decisions imposes strict runtime requirements on the neural networks that run on embedded hardware in an autonomous vehicle. By compressing the neural networks, for example using fewer parameters and bits, the algorithms can be executed faster and use less energy. For this task, the CERN–Zenseact team chose field-programmable gate arrays (FPGAs) as the hardware benchmark. Used at CERN for many years, especially for trigger readout electronics in the large LHC experiments, FPGAs are configurable integrated circuits that can execute complex decision-making algorithms in periods of microseconds. The main result of the FPGA experiment, says Petersson, was a practical demonstration that computer-vision tasks for automotive applications can be performed with high accuracy and short latency, even on a processing unit with limited computational resources. “The project clearly opens up for future directions of research. The developed workflows could be applied to many industries.”
The compression techniques in FPGAs elucidated by this project could also have a significant effect on “edge” computing, explains Maurizio Pierini of CERN: “Besides improving the trigger systems of ATLAS and CMS, future development of this research area could be used for on-site computation tasks, such as on portable devices, satellites, drones and obviously vehicles.”