Integrating model-based machine learning methods into deep neural
architectures allows one to leverage both the expressive power of deep neural
nets and the ability of model-based methods to incorporate domain-specific
knowledge. In particular, many works have employed the expectation