Continuous emotion recognition using a particle swarm optimized NARX neural network

The recognition of continuous dimensional emotion remains a challenging task due to large variations in the expression of emotion, and the difficulty of modeling emotion as temporal processes. This work proposes the use of a Nonlinear AutoRegressive with eXogenous inputs recurrent neural network (NARX-RNN) to learn emotional patterns in a given a dataset. The application of particle swarm optimisation in training the NARX-RNN is considered and compared to a gradient descent algorithm. We show that the NARX-RNN outperforms other methods in its emotion recognition ability, and can be easily trained with both gradient-free and gradient-based optimization methods.