© Hector Fellow Academy
8. March 2022
New publi­ca­tion by Maxim­il­ian Dax

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: