ImprovingVariationalAutoencoderswithDensity Gap-based Regularization