Supervisors: Dr. Zhiwu Huang, Dr. Danda Paudel, and Prof. Luc Van Gool
Emotion prediction is one of the important problems in modern AI. The desire to create machines that are able to interact with humans neccesitates ways of discerning human emotional state from the available signals. One of the most important is vision and recognition of emotion from a human face. The accuracy of predictions made with neural networks in this case, as well as in every other is tied to the quality of the provided labels. As these labels are provided by human experts, uncertainty is unavoidable. We analyze the quality of the labels provided by human experts and their significance in the ongoing research. Generative Adversarial Networks are one of the most popular topics in mod- ern AI research. They are based on specific architecture relying on the in- terplay between so called generator and discriminator that is described by GAN loss. We investigate the influence of inclusion of this loss in addition to the standard MSE on generalization of the neural network architectures.