Current techniques for recognizing human expressions and emotions rely heavily on the extraction and analysis of facial features. This approach tends to be reliable thanks to the high number of facial activations whose mixture allows a clear discretization of the different human states. Evidently, the temporal context and evolution of these activations has a not insignificant effect on their analysis, since the knowledge of a previous state may contain relevant information for the recognition of the next state. In this way, we propose a study of the recognition of human facial expressions using recurrent neural networks, capable of retaining temporal information and analyzing sequences of images instead of static images. As a work plan, you should initially collect existing databases, proceeding to the design of recurring structures and with different types of cells (e.g. LSTM, GRU, etc.). It is intended to compare the different architectures and create a robust classifier of facial expressions.