First exploration of the potential of diverse training and voting for increasing the accuracy of CNNs

Machine learning techniques are attracting a huge amount of interest from both industry and academia. For instance, Convolutional deep Neural Networks (CNNs) have recently enjoyed a notable success in image understanding. The automotive industry is already using image classifiers for Advanced Driver-Assistance Systems and in the development of the upcoming autonomous cars, which will have to guarantee high levels of reliability. The certification of systems based on machine learning is an open issue but it is clear that any improvement in the performance of image classifiers is to be welcomed. CNNs need to be trained to act as image classifiers. This training leads to slightly different classification capacity depending on some training parameters. In this paper we present a first exploration on the use of schemes based on voting on the results of several CNNs trained differently, as a means to increase the final classification performance, and thus the reliability, of this type of systems.