Fast and Unified Path Gradient Estimators for Normalizing Flows
This work addresses a bottleneck for researchers and practitioners using normalizing flows in variational inference and maximum likelihood training, offering a more practical solution, though it is incremental in improving existing estimators.
The paper tackled the computational inefficiency and scalability limitations of path gradient estimators for normalizing flows, proposing a fast and unified estimator that improves efficiency and applies to maximum likelihood training, with empirical results showing superior performance and reduced variance in natural sciences applications.
Recent work shows that path gradient estimators for normalizing flows have lower variance compared to standard estimators for variational inference, resulting in improved training. However, they are often prohibitively more expensive from a computational point of view and cannot be applied to maximum likelihood training in a scalable manner, which severely hinders their widespread adoption. In this work, we overcome these crucial limitations. Specifically, we propose a fast path gradient estimator which improves computational efficiency significantly and works for all normalizing flow architectures of practical relevance. We then show that this estimator can also be applied to maximum likelihood training for which it has a regularizing effect as it can take the form of a given target energy function into account. We empirically establish its superior performance and reduced variance for several natural sciences applications.