The Astrometry program first searches the image to identify a set of four stars (quad), and then searches the catalogue for similar quads (the green outline in the top right corner; in this case, it looks like a triangle because two of the stars are so close together). When it finds an apparent fit, it checks the other stars in the image (red circles) against the stars in the catalogue (green circles) until there's a match. (Astrometry.net)

Valentine's Day 2007 was especially sweet for the team of computer scientists and astronomers trying to build an "astronomy engine." After almost two years of work, they were able to take a starscape – a picture of the Rosette Nebula - and identify it using their computer program.

"We solved it - that is, automatically determined its location on the sky and orientation -right out of the box," they announced on their website.

The test validated their plan to build a computer system that could take any image of the heavens – digital and film camera pictures, from amateur telescopes, observatory telescopes or space telescopes like the Hubble – and determine the precise area that is pictured. In effect, Astrometry.net is mapping pixels and the sky.

The project, which is expected to be available through the web, will help amateurs and professionals alike. Astronomers love it, said Dustin Lang, a University of Toronto PhD student who worked on the programming. "Most modern telescopes are really complicated computer systems," and because they're computers, they have glitches, leaving astronomers trying to identify images.

Beyond that, astronomers may know what they're looking at, but want the exact location. And then there are an estimated three million old photographic plates with images of the night skies dating back to the 1880s.

Those plates would help modern research, allowing then-and-now comparisons of the skies, or providing evidence to support or knock down astronomical theories. But until the pictures are digitized – a project the International Astronomical Union and Harvard University, with about 500,000 plates, are both pushing – the data may as well not exist.

The plates are great pictures, but hard to annotate, Lang said, seeing another application for what the team calls "the solver."

And then there are the huge numbers of pictures taken by amateur astronomers. They often don't record exactly what they're shooting, but the program can find out. It's the amateur astronomer's equivalent of suddenly having a time and place appear on the back of all those old family snapshots.

Shining among the stars

New York University astronomer David Hogg is one of the key members of the Astrometry team. (Christopher Stumm/Astrometry.net)

The idea came from Sam Roweis, a rising U of T star, machine learning professor and Lang's PhD supervisor. Roweis went to high school in Toronto with New York University astronomer David Hogg, and their discussions led to a U.S. National Science Foundation grant to launch Astrometry.net.

Now Roweis, Hogg, Lang, former U of T student Keir Mierle and NYU astronomer Michael Blanton are the main drivers of the project. Roweis saw that identifying star pictures is similar to the problem computers face in trying to recognize objects. That is an aspect of machine learning, an area that deals with the ways computers recognize patterns, classify data and make predictions. So while he is at Google in the summer of 2007 – raising the possibility that the finder might eventually be available through the popular service – Roweis's core work includes applications like teaching computers to understand speech, read handwriting and recognize faces, the U of T website said.

University of Toronto machine-learning professor Sam Roweis.

Roweis, who holds a Canada Research Chair in statistical machine learning, was picked in June 2007 as one of 50 recipients of a special grant awarded by the Natural Sciences and Engineering Research Council. "Based on their success and accomplishments so far, we believe they are poised to make real breakthroughs in their fields," council president Suzanne Fortier said of the winners.

The finder is a complicated piece of engineering, but rendered simply, it operates like a web-page search. But rather than finding pages with "machine learning," the system tries to match stars in an image with a catalogue of known pictures of the sky.

There are two key challenges. With over a billion stars out there, matching a picture that may be less than one-millionth of the sky's area is daunting, Lang said.

Secondly, "both catalogues and pictures are noisy" in that one may contain extra stars that are not in the other, and some catalogue stars may be missing from the image, Roweis said in a presentation.

These "distracters" and "dropouts" hugely complicate finding a match.

The solution is to start small. When an image is presented to the computer, it looks at sets of four stars (quads) and searches for matches of those four in the catalogue. "Each time it finds a match, it asks, 'if this quad in the image matches that quad in the index, where would I expect to find other stars?" Lang said.

If the image contains stars in the places predicted by the catalogue, there's a match. That may take a few thousand tries, but 99.84 per cent of 336,554 images were identified correctly in one test. Only 530 baffled the system, and there were no false identifications.

In July 2007, astronomers are running trials with the finder, and the team's working on making the process faster, more flexible and more robust. "We also hope to insert the system into the observing pipeline of telescopes, debug standard catalogues, learn about individual instruments and facilitate 'collaborative observing' tools," Roweis said in his presentation.

Opening it to the public depends partly on when the team is ready to move into maintenance and user support mode, Lang said. That will probably happen in late 2007, but already, they're talking about expanding the information available. From the current version, which provides the location of the image and some interesting things it contains, he said they'd like to:

• Group amateur images to form a high-resolution picture of the sky that is constantly being updated, which could become a research tool for professional astronomers.
• Show other images (infrared or ultraviolet, for example) of the same part of the sky.
• Suggest other places to look. ("If you like this part of the sky, you'll love Orion's nebula".)

The sky is no limit, at all.

[an error occurred while processing this directive]