AI System Trained to Recognize New Galaxies

Researchers from the University of Western Australia have developed a deep learning system that can identify galaxies in deep space.

The system, called ClaRAN, presents a system that scans images taken by radio telescopes and spots radio galaxies that emit powerful radio jets from their black holes.

Dr. Ivy Wong, an astronomer from the University of Western Australia, and an author of the research says black holes are found at the center of most, if not all, galaxies.

“These supermassive black holes occasionally burp out jets that can be seen with a radio telescope,” Wong said. “Over time, the jets can stretch a long way from their host galaxies, making it difficult for traditional computer programs to figure out where the galaxy is. That’s what we’re trying to teach ClaRAN to do.”

The work uses the Faster Region-based Convolutional Neural Network (Faster R-CNN), developed by Microsoft and Facebook researchers as a base. The team says the program was completely overhauled and trained to recognize galaxies instead of people.

ClaRAN looks at over 500 different views of radio galaxy data to make its detections and classifications. After scanning through the different views, ClaRAN then also considers the data from infrared telescopes to refine its predictions, giving the final detection and classification result of a radio galaxy jet system. Credit: Dr Chen Wu and Dr Ivy Wong, ICRAR/UWA.

Using NVIDIA Tesla GPUs and the cuDNN-accelerated TensorFlow deep learning framework, the team trained their convolutional neural network on thousands of world coordinate system-aligned radio and infrared image pairs.  The neural network then classifies them into one of six morphology classes.

Dr. Wong says traditional computer algorithms identity around 90 percent of the sources. “That still leaves 10 percent or seven million ‘difficult’ galaxies that have to be eyeballed by a human due to the complexity of their extended structures,” Dr. Wong explained.

“If ClaRAN reduces the number of sources that require visual classification down to one percent, this means more time for our citizen scientists to spend looking at new types of galaxies,” she said.

By combining the data from different telescopes, ClaRAN’s ‘confidence’ level in its detections and classifications is increased. Shown as the number above the detection box, a confidence of 1.00 indicates ClaRAN is extremely confident that the source detected is a radio galaxy jet system and that is has classified it correctly. To the left is a radio galaxy jet system detected by ClaRAN using only data from radio telescopes. ClaRAN isn’t sure what it’s seeing here, giving two predictions, one covering the entire system with a low confidence of 0.53, and one covering the top jet only with a confidence of 0.67. To the right is the same galaxy, but with infrared telescope data overlaid. With the inclusion of data from infrared telescopes ClaRAN’s confidence in the detection has increased to the highest value of 1.0, and ClaRAN now includes the entire system in its only prediction. Credit: Dr Chen Wu and Dr Ivy Wong, ICRAR/UWA

The researchers have published the code for ClaRAN on GitHub.

A research paper describing the method was also published in Monthly Notices of the Royal Astronomical Society, published by Oxford University Press today.

Read more>


To see more amazing breakthroughs, make sure to attend NVIDIA Founder and CEO Jensen Huang’s keynote at SuperComputing 2018 in Dallas, Texas on November 12 at 3:00 PM CT