• Loss function
• Emotion classification
• Explainable artificial intelligence
• Informed machine learning

This thesis work proposes a new loss function, called LogRatio loss and divided into three parts (LogRatioFarthest, LogRatioNearest, LogRatioRandom). This new loss function was used, together with categorical crossentropy, to train an end-to-end architecture that takes in psd data and classifies categorical emotions. For performance comparisons, other architectures were also trained, in one-dimensional, two-dimensional and three-dimensional cases. In the thesis there is a large section where the results that have been obtained in the various experiments are reported.