On the Potential of Randomization-based Neural Networks for Driving Behavior Classification

Naturalistic Driving has recently garnered the attention from the community working on Deep Learning models, producing a plethora of modeling proposals on account of reaching an increasingly better predictive performance. Little attention has been paid to the computational implications of adopting such models, particularly when used for inferring the behavior of the driver from naturalistic driving data. This work enters this uncharted research area by probing the balance between complexity and performance of randomization-based neural networks for driving behavior classification. To this end, results from an extensive experimental benchmark comparing these networks to diverse Deep Learning and ensemble learning models are discussed, unveiling a significantly better-balanced trade-off between performance and complexity of randomization-based neural networks, and suggesting more concern with the efficiency of models in prospective studies.