Engineering news
Huge optical observatories and giant mushroom-like radio antennas now do the job. And to spot new events such as supernovas or pulsars, scientists use automated surveys to scan the sky day in and day out.
But here comes the problem. While such surveys find plenty of 'candidates', it then takes astronomers a lot of time to sift through the data and filter out events that don’t look promising. Given the huge volume of data available today, it has become impossible to do manually - and that's where machine learning comes in, as an efficient method to analyse large data sets obtained by modern telescopes.
This type of machine learning is not just pure science. It can, and already is, being applied outside of astronomy – for example in medicine and manufacturing, said Rob Lyon, a computer scientist who specialises in machine learning at the School of Physics and Astronomy at Manchester University. He spoke to PE at a conference in Jodrell Bank, just outside Manchester, where astronomers and telescope engineers have gathered to celebrate 50 years since Jocelyn Bell Burnell discovered the very first pulsar.
"Machine learning can be used for control systems that need continuous monitoring, and you need to get a prediction [of] when something may break," said Lyon. "It can also be used for optimisation problems, for an engineering challenge which requires a selection of parameters which you're unsure of. Machine learning methods can help you do that.
"I have a colleague who's using it as part of an analytical model for heart rate monitoring, predicting when the human heart will undergo some kind of an event, for instance a heart attack."
So how does this kind of machine learning work, for instance, in astronomy?
First, the basic computer code used to search for pulsars produces diagnostic plots. Traditionally, people would then look at those plots and determine what is and isn't a promising candidate, said Ryan Lynch, an astronomer at National Radio Astronomy Observatory who uses machine learning to find pulsars.
"The machine learning code we use in our survey is a deep neural net that analyses the image data that goes into those plots to look for patterns that seem pulsar-like. The neural nets require a training data set that has already been looked at by astronomers - it then tries to get the same answer and repeats that process until it becomes good enough."
Astronomers started using algorithms based on neural networks in 2010, and since then, the process is getting ever more sophisticated, said Lyon.
Another astronomer who actively uses machine learning is Chia-Min Tan, a postgraduate student at Manchester University, who works on the LOFAR Tied Array All-Sky Survey (LOTAAS), which scans all northern sky surveys for pulsars and fast transients.
First, she and her team extract numerical information as input for a machine learning classifier to distinguish between pulsars and non-pulsars. "The machine learning classifier is built by selecting a sample of examples of pulsars and non-pulsars from our survey data and extracting the same features from them for a machine learning algorithm (in our case a decision tree algorithm) to figure out how to classify the data set that we have given it," said Tan.
"The decision tree algorithm works by looking at the features and attempt to find the best feature to separate the data that was given into branches. The process will repeat in each branch until we obtain a classifier that gives us an ideal performance."
Robots are unlikely to entirely replace astronomers one day, though, said Lynch. "I do think the tasks that can be described in some sort of standard recipe or algorithm will become increasingly automated," he said. "But astronomers will still be required to interpret the results and come up with new ideas."