Fall and Winter is a good time to look for the Andromeda galaxy in the Northern Hemisphere. Find a dark area on a moonless night and you should be able to easily find it. At about two and a half million light years away it's the most distant object you can see with an unaided eye.1 Twice as big as our own Milky Way and nearly a trillion stars, it takes up six times as much sky as the full moon. The beautiful photos you see of it are usually several images stitched together to cover it's full extent. And that illustrates a problem astronomers have had.
Most telescopes look at very small regions of the sky, but a lot of interesting things happen almost randomly and the pros often come upon them later. For about one hundred and fifty years astronomers have had a friendly relation with amateur astronomers to get wind of something new events as well as working with data. When you study something so big, you are grateful for enthusiastic help.2
In the past ten years astronomy has seen an enormous increase in data to study. A few large scale sky surveys are recording far more information than the tiny community can study. In the near future the Large Synoptic Survey Telescope comes on line looking at over 37 billion stars and galaxies every night making a huge video of part of the sky that will continue for decades. Several other projects are coming on line that generate data on the order of what comes out of CERN. Astronomy is likely to go through another revolution, but while there have been some citizen-science projects have helped with real discoveries, the fundamental problem is new huge data firehoses are growing at a fast clip. Experienced people don't scale.
The Dark Energy Survey has taken a crack at the problem. It had a few dozen scientists looking carefully for about a half year at images from an area of the sky that covers about the same amount of area as a thousand full moons. They were looking for warped shapes that mark strong gravitational lensing.3 About a thousand of these objects are known, but a very small number are lensed by supernovae and happen to be extremely interesting. For that reason the astronomers don't want to miss anything.
Machine Learning is good at recognizing shapes and it seemed to be the reasonable approach. Once the scan had been done it was decided to see if larger areas could use neural networks. The DeepLensing project's neural network is still under development . At this point machine learning picks up about 90% of the known objects. It may miss some important objects, but it's about a million times faster than a person, so even with ten thousand good volunteers it will cover more sky and find more objects.
Neural nets are being used to find odd-ball objects ... one project is looking at data from the now defunct Kepler spacecraft and organizing objects by "weirdness"- the mark of potential discoveries. The idea is to flag unusual objects and events for further study. So far not much has been found, but a few existing very strange stars that weren't in the training data have been spotted.
My favorite so far is the Robotic Exoplanet Recognition program. RobERt studies spectrums from the bit of light that makes it through an exoplanet's atmosphere. There's not much data and it isn't terribly complex. The problem is the hammer and tongs approach of using theoretical atmosphere models of these atmospheres that might match the observed spectral images is a difficult computational task. It can take a week to get a result to tell the a space based telescope to have another look. That's way too long. RobERt was trained on what water, carbon dioxide and TiO in the spectrum look like and makes a rapid determination.
Now the fun part. There was a good deal of experimentation playing with the size and design of the network. Once it was working it was decided to run it in reverse. Using the output side as the input the label 'water' was entered. This would be like using our number recognizer after it had been trained by putting 7 as an input and seeing what the new output - the 784 neuron first layer - would look like. Ideally it would be something like a hand-drawn 7.
At first the water trial failed. It turned out there were far too many neurons that weren't doing much other than making noise. The network was redesigned with very few neurons - just enough to do a good job on regular classification. Now when the input and output were reversed entering the label "water" made a plausible spectrum. No need to really understand it - just look at the similarities. This helps people think about interesting atmospheres and that can lead to insight.
The reversal of input and output is referred to as dreaming. The machine was dreaming novel potential spectrums for water (also carbon dioxide and titanium oxide)
There are many flavors of machine learning and "AI". Several other experiments are underway, but astronomy is a physical science and you're supposed to know deep things about data processing and the provenance of your data. You only use machine learning when you have to and you need to approach it with great caution. Even so there is a huge need and enormous potential for improvement. I wouldn't be surprised to see fundamental improvement come from machine learning users in fundamental research rather than industrial computer science.
__________
1 With the naked eye you mostly see the core of the galaxy - use binoculars to reveal a much larger extent.
2 Some areas of observational biology - traditional field work - makes use of amateurs. Ornithology in particular. Cornell's eBird program aimed at tapping birders worldwide as part of a amateur science extension. Participation was light until they realized they could make the program attractive to birders. Some changes were many and now eBird has over 300,000 participants many of whom are heavily invested in the community,. By making a useful birding program they've managed to create the largest source of good data ornithologists have ever dealt with and it's in usable digital form. Big data.
3 General Relativity tells us light traveling past a mass like a star gets bent. Unseen masses can be detected by looked for warped shapes.
__________
Comments