The paper of doctoral researcher Maximilian Dax and his colleagues will be presented at conference
In their paper, the scientists describe a new method for solving inverse problems of science.
While so far there are only simulation-based inferences with conditional neural density estimators that treat the underlying forward model as a black box without exploiting geometric properties such as equivariances, in their paper the researchers present an alternative method for including equivariances in joint transformations of parameters and data.
Equivariances are common in scientific models.
The alternative method, called group equivariant neural posterior estimation (GNPE), applies to both exact and approximate equivariances. As a real-world application, the GNPE research team uses amortized inference of astrophysical binary black hole systems from gravitational wave observations. The GNPE achieves state-of-the-art accuracy and can reduce inference times by three orders of magnitude.
The paper was accepted for a poster presentation at the 10th International Conference on Learning Representations (ICLR). The ICLR is the first conference dedicated to advancing the area of artificial intelligence known as representation learning, or more commonly, deep learning.
The authors have made the paper available on the pre-print server arXiv: