Effects of Modifying the Input Features and the Loss Function on Improving Emotion Classification

In this work, we show that the discriminative power of a deep neural network can be improved at three different levels: (i) inputting discriminative features, (ii) designing an optimal network architecture and (iii) changing the loss function. This work suggests that only increasing the depth or the width of a deep network may not always be the best solution while requiring computationally efficient models. Here, we show that there is scope for improving the classifier accuracy at each of the three levels. We have carried out all our experiments on the FERplus dataset and show that the facial emotion recognition accuracy can be independently improved up to 2.5% by adding better features, 1.8% by modifying the loss function and up to 3.1% by combining the two ideas. In separate experiments, we show that the computational complexity can be reduced by a factor of 24.3, while simultaneously increasing the FER by 0.95% by modifying the architecture of the model, input features and the loss function.