A Deep Learning Architecture for Egocentric Time-to-Saccade Prediction using Weibull Mixture-Models and Historic Priors