Scientists at Stanford University and SLAC National Accelerator Laboratory have demonstrated that neural networks can accurately analyze gravitational lenses (complex distortion in spacetime) about 10 million times faster than conventional techniques.
Today’s methods require input from experts and are computationally demanding. They usually take weeks or even months to complete. However, the same results can be achieved by neural nets within a fraction of a second, in a completely automated way, using less computational resources.
Superfast, Automated Complex Analysis
The research team used neural networks (a form of AI) to examine the pictures of strong gravitational lensing. This process involves multiplying and distorting the photo of distant galaxy into rings and arcs by the giant object’s gravity like galaxy cluster, that is near to us. The distortions give crucial details about the mass distribution in space and how it changes over time – data related to invisible dark matter and dark energy responsible for expansion of the universe.
This is a monotonous process of comparing a huge volume of machine simulations with actual photos, which could take weeks or months (for a single lens).
However, with the help of neural networks, the accurate analysis could be performed within seconds. The researchers have shown this using actual pictures (taken from Hubble Space Telescope) and simulated images.
Hubble Space Telescope images
For testing purpose, neural network is trained on around half a million simulated pictures of gravitational lenses. After training, network were quickly able to analyze new lenses with a accuracy rate comparable to traditional approach. It also determines the uncertainties in analyses.
‘3 publicly available neural nets and one developed by the researchers, were tested to deduce the properties of all lenses, which includes analyzing how the mass was distributed and how much it amplified the picture of the background galaxy’, says Yashar Hezaveh, lead author of the research.
This is much more advanced as compared to recent neural network applications in astrophysics that only address classification problems like figuring out whether pictures correspond to a gravitational lens or not.
The ability to look through and analyze a large volume of complex data in a short period of time, in a completely automated way is very much required for future space surveys – for looking deeper into the space and generating more information than ever before.
The LSST (large synoptic survey telescope) having 3.2 gigapixel camera is currently being developed at SLAC. It will help us to explore unparalleled views of the universe and provide us far more strong gravitational lenses, from a few tons to tens of thousands.
‘In future, we probably won’t have enough engineers to analyze this much of data with conventional approach‘ said Perreault Levasseur, coauthor of the study. ‘AI will help us identify and analyze space-objects quickly, giving us more time to focus on meaningful question about our universe’ he added.
A Revolutionary Method
A neuron is a single computation unit corresponds to a image pixel that is being examined. The neurons are generally organized into hundreds of layers, with each layer searching for the features present in the image. When the first layer discovers a specific feature, it passes the data to the next layer, which then tries to find another feature within that feature, and this process goes on.
‘What’s really interesting is neural networks learns themselves after you fed a bunch of data. They know what to look for in next phase‘ said Phil Marshall, coauthor of the paper. ‘It’s like small children learning to recognize objects. No one tells them what exactly a cat is, instead one shows them images’. he added.
However, in the case of neural network, it picks images of cats from a bunch of pictures, and returns some additional information like cat’s height, color, age, and weight.
The new neural network algorithms merged with modern central or graphical processing units, can provide reliable outcomes in less time, as the gravitational lens issue addressed in this research. In future, the same approach could be used for solving complex problems in other fields also.