Unsupervised Learning of View-invariant Action Representations