implicit_vi.ipynb - Colaboratory
colab.research.google.com · 4 words · saved by 1 readers
Reparameterization allows us to use gradient optimization to fit a proposal distribution 𝑞 𝜃 (𝑧) to the posterior by optimizing the ELBO. In this lecture you will see how to achieve the same goal but with using a richer family of possible 𝑞 𝜃 (𝑧) . You will be introduced to these families of models, and be given enough time to try to understand how we can compute gradients and optimize the parameters.