Flow-Transformed Implicit Processes for Function-Space Variational Inference
Flow-Transformed Implicit Processes (FTIP) is a method for function-space inference that uses a normalizing flow to create a richer variational distribution over surrogate variables. This allows it to model complex, asymmetric, and multimodal posterior distributions of functions. FTIP combines the flexibility of implicit priors with the expressive power of normalizing flows while maintaining scalable training methods.