Exclusive: Leading astrophysics researchers explain how they trained a neural network to classify images of galaxies
Artificial Intelligence (AI) has reached yet another milestone, as a group of scientists announce that they have successfully trained a convolutional neural network to analyze and classify images of galaxies.
This is set to supercharge our breadth and depth of understanding of some of the most fundamental questions of the universe; how do galaxies form and evolve?
In a study set to be published in The Astrophysical Journal, a team of scientists from the University Paris Diderot and UC Santa Cruz used computer simulations of galaxy formation to train a deep learning algorithm on how to classify real images of galaxies from the Hubble Space Telescope.
So far, images of galaxies have been categorized manually by human astronomers and astrophysicists; a time-consuming process that often involves many years, resources, and efforts.
The latest advances in deep learning, however, could soon make these practices obsolete, significantly speeding up research investigating the formation and evolution of galaxies.
The study recently carried out at UC Santa Cruz stems from years of scientific collaboration and is based on established theories of galaxy formation.
Clique has spoken to two researchers who participated in the research project, diving deep into the fascinating story behind this recent deep learning application.
AI to understand the evolution of galaxies
The idea of using deep learning algorithms to classify galaxies dates back to 2014, specifically to a project carried out by Sander Dieleman, who was a graduate student at the University of Ghent, in Belgium.
At the time, the Galaxy Zoo project had asked almost a million volunteers to help astronomers classify images of galaxies by answering a series of questions about them. Using a convolutional neural net, Dieleman achieved a success rate of 99% in predicting what these image classifications had been.
Dieleman now works at Google DeepMind in London and was one of the members of the team that wrote the code for the company’s world-famous AI, AlphaGo.
His successful graduate project was one of the key inspirations behind the study carried out by the researchers at UC-Santa Cruz.
Marc Huertas-Company, first author of the recent study, asked Dieleman to borrow the code he developed, modify it, and use it for further research.
This eventually led to the collaboration with the other scientists at UCSanta Cruz, where Huertas-Company spent the past two summers.
Said the paper’s co-author Joel Primack, professor emeritus of physics at UC- Santa Cruz and is renowned in his own right;
“We started working closely with Marc in the summer of 2016. Marc started by taking about 8000 galaxies that a team of 65 astronomers from the CANDELS project had classified by eye, and using them to train a deep learning code.
He showed that the neural net could classify galaxies as well as any astronomer, but also that machines are much faster, as just in a couple of days he classified images that astronomers had analyzed in over three years.”
Neural networks to classify galaxies
In their recent study, Huertas-Company, Primack, and their colleagues trained their convolutional neural network on output from computer simulations of galaxy formation.
When human astrophysicists observe galaxies in the universe, they truly only get a few snapshots of these, making it difficult to investigate their formation and evolution over time.
Galaxies are, of course, almost incomprehensibly large, stretching hundreds of thousands light years across, which basically means that it takes over 100,000 years for even light to move across them.
In order to deduce more about their stages of evolution, therefore, astronomy researchers have to look at how images change and compare them to simulations. If simulated images and real observations are similar, this would suggest that their theories are on the right track.
The simulations used by the researchers at UC- Santa Cruz were created by Primack and other international researchers, and identified three main stages of galaxies’ evolution: the Pre-Blue-Nugget, Blue-Nugget, and Post-Blue Nugget phase.
Throughout these three stages, which last billions of years, galaxies change in shape and appearance, as gas flows into the center of a galaxy and forms a small region with high density of stars, dubbed ‘blue nugget’.
By analyzing both simulations and observations, the neural network was able to successfully classify real images of galaxies, also suggesting that the blue nugget phase only happened in galaxies that had masses within a particular range.
“Initially, we were very suspicious that the code was cheating, in the sense that it was actually trying to estimate the mass of the galaxy and it was really classifying the galaxy based on its mass.
In that case, ours would have been a backward conclusion, that was not proving that the simulations and the observations were similar but that was actually using an external feature.”
To test whether the algorithm was working properly, the researchers went through the whole process again, but making some galaxies artificially brighter and others dimmer, to prevent it from cheating.
“We got exactly the same result, which that was really impressive,” said Primack. “I was really pleased, because it meant the code was doing the right thing; looking at the shapes of the galaxies and the distribution of light.”
Promising tech for astronomy research
The recent research carried out at UC- Santa Cruz revealed how effective AI can be in analyzing and classifying images of galaxies, but the journey towards a widespread application of these tools has merely begun, Primack cautiously explains;
“We don’t want to make too strong a claim. So far, we resolved about 35 galaxies, which is very small when compared to the hundreds of thousands of galaxies that we actually have data on from Hubble Space Telescope and other telescopes.”
The study was a first, very successful attempt at training convolutional neural networks on high quality simulations, to then use them to classify images of galaxies into different evolutionary stages.
The researchers are now planning to train their algorithm on more simulations and use it to explore other stages of galaxy formation, as well as further aspects of their evolution.
Primack confirms that there is much to be done;
“I think that this is the beginning of a very long road. I’m sure our competitors in the field are going to do similar things, because this approach to deep learning and convolutional neural nets is very popular and many people are learning it now.
We ourselves have many more simulations of different galaxies and we’re going to be doing a wider variety of studies. I’m very happy that we were the first to carry such a thorough project out, actually showing that it seems to work.”
Over the past few years, deep learning algorithms have showed great promise for astrophysics research, and they could soon play a major role in finding answers to these profound questions.
David Koo, professor emeritus of astronomy and co-investigator in the UC-Santa Cruz study, said:
“AI will definitely be a growing aspect of doing astrophysics, just for handling the explosive amount and richness of data from both theoretical computer simulations and from extremely powerful new telescopes soon to be operational or in planning stages.
This role of AI will be important for all areas of astrophysics, not just galaxies. Studies of stars, planets, black holes, etc. are all exploding in the amount and richness of data.”
According to Koo, images from space are set to grow exponentially – easily resulting in 100 to 1,000 times more available data than today. AI could help to delve into this data quicker, resulting in ground-breaking discoveries in far shorter periods of time.
“A bit tongue in cheek, AI may also help to expand the pool of people doing astrophysics, bypassing the need for years of technical training to learn the skills and gain the expertise needed to do the nitty gritty parts today of scientific analysis.
“AI could make astrophysics research much more democratic…moving citizen science to a whole new level. I can imagine astrophysics research no longer resembling most of what we do today.”