Variational inference with NoFAS: Normalizing flow with adaptive surrogate for computationally expensive models

Fast inference of numerical model parameters from data is an important prerequisite to generate predictive models for a wide range applications. Use sampling-based approaches such as Markov chain Monte Carlo may become intractable when each likelihood evaluation computationally expensive. New combining variational with normalizing flow are characterized by computational cost that grows only linearly the dimensionality latent variable space, and rely on gradient-based optimization instead sampling, providing more efficient approach Bayesian about parameters. Moreover, frequently evaluating expensive can be mitigated replacing true offline trained surrogate model, neural networks. However, this might significant bias insufficiently accurate around posterior modes. To reduce without sacrificing inferential accuracy, we propose Normalizing Flow Adaptive Surrogate (NoFAS), strategy alternatively updates We also sample weighting scheme training preserves global accuracy while effectively capturing high density regions. demonstrate superiority NoFAS against various benchmarks, including cases where underlying lacks identifiability. The source code experiments used study available at https://github.com/cedricwangyu/NoFAS.