Check out this terrific trailer for our upcoming NIPS abstract (with Alex Wiltschko, Matt Johnson, Ryan Adams and David Duvenaud) put together by David – it does a really nice job explaining the value of merging a model-based approach with neural nets. The combination allows clear articulation of model structure – and maintenance of semantic meaning – while simultaneously taking advantage of flexibly learned feature embeddings. We think this is going to be an important and general method for capturing structure in high-dimensional data. If you are at NIPS this year, check it out!