Technology: Neural network takes the twinkle out of stars

日期:2019-03-03 08:14:01 作者:屠梓荛 阅读:

By ELISABETH GEAKE Next week a group of astronomers and technologists hopes to test a new method for taking the twinkle out of stars. They will be looking at the star Vega, using a ‘neural network’ computer program to adapt the shape of a mirror to help compensate for the way the atmosphere distorts starlight. Turbulence in the air continually changes its refractive index – its ability to bend light – making faint and distant objects appear to shimmer. This makes it impossible for astronomers to get accurate images of faint stars with telescopes on Earth. But the distortion can be cancelled using deformable mirrors, which bend the starlight back to where it should have been, and then record the image. Calculating nearly instantaneously the movements of such a mirror normally requires a powerful computer. The new method, reported this week in the journal Nature, uses an ordinary desktop computer running a neural network program instead. Neural networks model an array of processors and connections in software, based on the arrangement of neurons in the brain. By reinforcing certain connections, the network can be optimised to choose the correct mirror movements. The results from neural networks are sometimes not completely exact, but what is lost in precision is made up for by a cut in the amount of computation needed. Ed Zuiderwijk, who works on the La Palma telescope project, says: ‘You don’t want exact answers – you would use a linear program for that – but one that is good enough.’ Chris Dainty, professor of applied optics at Imperial College in London, adds: ‘Mathematically there are almost certainly better methods, but (one must consider) what is practical.’ David Sandler of Thermo Electron Technologies in California had the idea of using a neural network, and is currently trying to improve the technique. The system will be tested at the Multiple Mirror Telescope in Arizona,