Flexible Prior Distributions for Deep Generative Models
This work addresses a bottleneck in deep generative modeling for researchers and practitioners by improving model flexibility and interpretability, though it appears incremental as it builds on existing paradigms.
The paper tackles the problem of training deep generative models by proposing flexible prior distributions for latent codes instead of simple priors, demonstrating that this approach yields more powerful generative models, better latent structure modeling, and explicit control over generalization.
We consider the problem of training generative models with deep neural networks as generators, i.e. to map latent codes to data points. Whereas the dominant paradigm combines simple priors over codes with complex deterministic models, we argue that it might be advantageous to use more flexible code distributions. We demonstrate how these distributions can be induced directly from the data. The benefits include: more powerful generative models, better modeling of latent structure and explicit control of the degree of generalization.