Throwing the Safety Net: Reliable Machine Rear

Tübingen, Potsdam – When two black holes merge, they emit gravitational waves that travel through space and time at the speed of light. When these reach Earth, the signals can be detected by large detectors in the United States (LIGO), Italy (Virgo), and Japan (KAGRA). By comparing it to theoretical predictions, scientists can determine the black hole’s properties: mass, spin, orientation, position in the sky, and distance from Earth.

The Empirical Reasoning Division of the Max Planck Institute for Intelligent Systems (MPI-IS) in Tübingen and the Astrophysics and Cosmological Relativity of the Max Planck Institute for Gravitational Physics (Albert Einstein Institute/AEI) in Potsdam A team of researchers in the department has now developed a self-checking deep learning system that extracts information from gravitational wave data with great accuracy. Along the way, the system checks its own predictions about black hole coalescence parameters. This is a deep neural network with a safety net. A set of 42 gravitational waves detected from black hole mergers was successfully analyzed by the algorithm. This study was published in the journal on April 26, 2023. Physical review letter.

DINGO: Deep Neural Networks for Gravitational Wave Analysis

Researchers developed a deep neural network called DINGO (Deep Inference for Gravitational Wave Observations) to analyze the data. DINGO is trained to extract (or infer) gravitational wave source parameters from detector data. There was a press release about this in December 2021. The network learned to interpret real (observed) gravitational wave data after training with millions of simulated signals in various configurations.

trust but confirm

However, it is not possible to tell at first glance whether a deep neural network is reading the information correctly. In fact, one of the shortcomings of common deep learning systems is that incorrect results sound plausible. Therefore, MPI-IS and AEI researchers have added controls to their algorithms. Maximilian Dax, PhD student in the Empirical Reasoning Division at MPI-IS and lead author of the publication, explains: First, the algorithm calculates black hole properties from the measured gravitational wave signals. Based on these calculated parameters, gravitational waves are modeled and compared to the originally observed signal. Therefore, deep neural networks can cross-check their own results and correct them if in doubt. ”

Since the algorithm controls itself, it is much more reliable than previous machine learning methods. But that’s not all. “I was surprised to find that algorithms were often able to identify unusual events—real-world data that did not match the theoretical model. You can use it,” says Stephen Green, co-lead author and former senior scientist at AEI (now the University of Nottingham).

“While we can vouch for the accuracy of our machine learning method, this is almost unheard of in the realm of deep learning. Therefore, it is compelling for the scientific community to use this algorithm to analyze gravitational wave data.” said Alessandra Buonanno, author and director of the AEI’s Astrophysics and Cosmological Relativity Division. Scientists around the world are studying gravitational waves in large collaborations such as the LIGO Scientific Collaboration (LSC), an organization of more than 1,500 researchers.

Bernhard Schölkopf, director of MPI-IS, adds: It confirms the correctness of the “black box” neural network approach. ”

LSC membership

Maximilian Dax and Jonas Wildberger, two Ph.D. MPI-IS students, are also members of the LIGO Scientific Collaboration. Through this membership, they will be able to quickly access gravitational wave detector data and collaborate with relevant working groups. Their goal is to develop DINGO into a standard method for gravitational wave data analysis.