VDSM: Unsupervised Video Disentanglement with State-Space Modeling and Deep Mixtures of Experts

Disentangled representations support a range of downstream tasks including\ncausal reasoning, generative modeling, and fair machine learning.\nUnfortunately, disentanglement has been shown to be impossible without the\nincorporation of supervision or inductive bias. Given that supervision is often\nexpensive or infeasible to acquire, we choose to incorporate structural\ninductive bias and present an unsupervised, deep State-Space-Model for Video\nDisentanglement (VDSM). The model disentangles latent time-varying and dynamic\nfactors via the incorporation of hierarchical structure with a dynamic prior\nand a Mixture of Experts decoder. VDSM learns separate disentangled\nrepresentations for the identity of the object or person in the video, and for\nthe action being performed. We evaluate VDSM across a range of qualitative and\nquantitative tasks including identity and dynamics transfer, sequence\ngeneration, Fr\\'echet Inception Distance, and factor classification. VDSM\nprovides state-of-the-art performance and exceeds adversarial methods, even\nwhen the methods use additional supervision.\n