Self-check­ing algorithm inter­prets gravi­ta­tional-wave data

Together with an inter­dis­ci­pli­nary team from the Max Planck Insti­tute for Intel­li­gent Systems and the Max Planck Insti­tute for Gravi­ta­tional Physics, includ­ing Hector Fellow Bernhard Schölkopf, Young Researcher Maxim­il­ian Dax published a paper on Neural Impor­tance Sampling for Rapid and Reliable Gravi­ta­tional-Wave Infer­ence in the journal Physi­cal Review Letters.

In the publi­ca­tion, he devel­oped an algorithm that immedi­ately checks its own calcu­la­tions of merging black holes’ proper­ties and corrects its result if neces­sary – inexpen­sively and rapidly. The machine learn­ing method provides very accurate infor­ma­tion about the observed gravi­ta­tional waves and will be ready for use when the global network of gravi­ta­tional-wave detec­tors starts its next observ­ing run in May.

The devel­oped self-check­ing deep learn­ing system extracts infor­ma­tion from gravi­ta­tional-wave data very accurately. In the process, the system checks its own predic­tions about the parame­ters of merging black holes – a deep neural network with a safety net. A set of 42 detected gravi­ta­tional waves from merging black holes were success­fully analyzed by the algorithm: When cross-checked against compu­ta­tion­ally expen­sive standard algorithms, the results were indistinguishable.