Latent Space Refinement for Deep Generative Models
This work addresses a common problem in deep generative models for science and industry, offering a method to enhance density estimation, though it appears incremental as it extends existing ideas to all generative model types.
The paper tackles the challenge of achieving precise data probability density representations in deep generative models by proposing a Latent Space Refinement (LaSeR) protocol that uses iterated generative modeling to circumvent topological obstructions and improve precision, demonstrated on combinations of Normalizing Flows and Generative Adversarial Networks.
Deep generative models are becoming widely used across science and industry for a variety of purposes. A common challenge is achieving a precise implicit or explicit representation of the data probability density. Recent proposals have suggested using classifier weights to refine the learned density of deep generative models. We extend this idea to all types of generative models and show how latent space refinement via iterated generative modeling can circumvent topological obstructions and improve precision. This methodology also applies to cases were the target model is non-differentiable and has many internal latent dimensions which must be marginalized over before refinement. We demonstrate our Latent Space Refinement (LaSeR) protocol on a variety of examples, focusing on the combinations of Normalizing Flows and Generative Adversarial Networks.