The aim of the present work is to investigate the performance of an Ensemble of
Deep Convolutional Neural Networks for automated Facial Expression Recognition.
Emotion classification of facial images could be integrated in diagnostic systems
or exploited in several other fields, ranging from Human-Computer Interaction to
Data Analytics. Recent works have proven that convolutional neural networks are
suitable for features extraction and inference, and that ensemble voting guarantees
a significant boost in performance. We investigate whether pre-training individual
networks on different datasets contributes to differentiate their training procedures
with the objective to produce a more accurate ensemble of neural networks. The
experiments have been carried out using TensorFlow machine learning library and
exploited GPU in order to speed up computation. They show that in the proposed
experimental scenario, the Pretrain Strategy is not appropriate: it does not improve
significantly the ensemble accuracy and it is more expensive in terms of time and
additional data than other differentiation strategies.