Publication date: 
For the first time, astronomers have used an artificial intelligence method called active deep learning. Scientists from the Faculty of Information Technology of the Czech Technical University with the active participation of students and in cooperation with the Astronomical Institute of the Academy of Sciences used the potential of artificial intelligence for the first time in the history of astronomical research.

The article was published by the prestigious journal Astronomy and Astrophycics.

The discovery is due to a team of scientists from FIT Petr Škoda, Pavel Tvrdík and doctoral student Ondřej Podsztavek. They dealt with the solution within a large grant of the Ministry of Education, Youth and Sports called RCI. The active deep learning method is based on the interactive improvement of the predictions of a multilayer convolutional neural network based on the expert's opinion. Unlike commonly used procedures, the network itself will ask for advice in those cases where it is the least sure. The Czech researcher successfully applied the method to four million spectra from the world's largest archive of spectra acquired by the Chinese LAMOST telescope and discovered almost a thousand hitherto undescribed very rare space objects with emission spectral lines.  

"Although we extensively researched the astronomical literature and asked colleagues around the world, we seem to be the first to use active learning in conjunction with deep neural networks in astronomy," says Petr Škoda, CSc., one of the members of the research team. .

He believes that active deep learning heralds the future direction of the use of artificial intelligence in astronomy and other sciences.

"Two years ago, we put together a small astroinformatics team as part of the RCI project. The mentioned publication is the first important output of our astroinformatics team, "adds prof. Pavel Tvrdík, co-author of the thesis and leader of the research group at FIT.

Researchers at FIT have other ideas for the future, specifically the connection of deep neural networks with large astronomical data and with methods not very commonly used in natural sciences. They want to focus on the active involvement of man in machine learning.