Unsupervised learning applied to progressive compression of time-dependent geometry