Generative Models using Data Representations¶
Preface¶
We will discuss how the data representation approaches allow us to learn more complex generative models, not just complex predictive models.
If we only stuck to simple generative models: Simple Generative Models, Mixture Models, then our generative models wouldn't be able to learn truly complex distributions.
First, we talk about increasing the capacity of a model by mapping to a new space with a data representation, and then using simpler parametric models.
Secondly, we talk about directly learning complex distributions with neural networks, using the idea of reparameterization. See The Reparemeterization Trick in Machine Learning
Connection of Generative Models to PPCA¶
Connection of Generative Models to PPCA
Variation Autoencoders¶
References¶
Principal Component Analysis (PCA) Probabilistic Principal Component Analysis (PPCA) Generative Models