Driving Speed as a Hidden Factor Behind Distribution Shift
Name
Kristjan Roosild
Abstract
This work investigates the effects of using a self-driving neural network on data generated at a different driving speed, e.g. when testing a model at low speed for safety reasons. Using Donkey Car, an autonomous toy car, we find that weak performance at deployment may indeed be caused by driving at an unsuitable speed. Driving at a novel speed makes the system behave worse via two mechanisms. Firstly, a change in appropriate outputs (e.g. turning angle) happens, as faster driving requires more aggressive turning. Secondly, for models using multiple consecutive frames as input, different speed results in out-of- distribution inputs, which are known to be troublesome for networks.
These findings show the importance of testing and deploying self-driving systems by using the speed known to them.
These findings show the importance of testing and deploying self-driving systems by using the speed known to them.
Graduation Thesis language
English
Graduation Thesis type
Master - Data Science
Supervisor(s)
Ardi Tampuu
Defence year
2022