Tuesday, 14 July 2009

Brains In Space

Or maybe "Space in Brains" might be closer.

Neural networks - computing devices whose basic structure is similar to biological neurology - have many applications. But we're still finding new ones.

One of the latest is using a brain-type computer to understand the Universe's gross structure.

From arXiv.org Astrophysics > Cosmology and Extragalactic Astrophysics comes Modelling the dusty universe I: Introducing the artificial neural network and first applications to luminosity and colour distributions by Alemida et al:
We introduce a new technique based on artificial neural networks which allows us to make accurate predictions for the spectral energy distributions (SEDs) of large samples of galaxies, at wavelengths ranging from the far-ultra-violet to the sub-millimetre and radio.
The Executive Summary version is at the ABC :
A technique based on how brain neurons behave could dramatically speed up computer simulations of the universe, say UK researchers.

Cosmologists have long used computers to simulate what the universe looks like and how it evolved. But modelling the universe in detail is an extremely time consuming process.

Therefore researchers, led by postgraduate student Cesario Almeida from the University of Durham, have used an artificial neural network (ANN) to speed up the process of creating mock catalogues of galaxies.

Their findings appear on the arXiv physics website.

ANN simulates the way brain neurons connect and compute information, and it can be used to solve a range of astronomy, mathematics and engineering problems.

By comparing mock galactic catalogues with actual observations, such as the ATLAS sky survey currently being conducted by the Herschel space observatory, cosmologists can assess how well their models perform.
Almeida and colleagues used the ANN to create their mock galactic catalogue in several wavelengths of light.

Overall, they found the ANN-derived universe matched previous, well-known models with an accuracy of between 80% and 90%. It was best at simulating galaxies in the near-infrared part of the spectrum.

"At all the wavelengths considered we find that the luminosity functions predicted by the ANN are in excellent agreement," they write.

Associate Professor Andrew Hopkins of the Anglo-Australian Observatory in Sydney, says the technique could speed up research in the field of cosmology.

"It's certainly exciting that it can speed up this process," he says. "It opens up a new approach to try and attack these problems because it can work very quickly on a large number of simulated galaxies."

"It's an important cog in a very large wheel and sidesteps a lot of messy detail."
Basically, we use our knowledge, our "best guess" of the way things work, to make predictions of what the Universe should look like. Making the predictions is really difficult, and until now, generating a model has been just too hard, or rather, would take too long.

This neural network does it with far fewer resources.

We can then compare our prediction with observations, checking our model against reality. The interesting bits are where they don't match. We can use those to figure out an even better guess, make a new model, check that against observations, and repeat the process, refining our knowledge over time. We must be pretty close already, as the initial models show an 80-90% agreement.

Before now, making a model was the hard bit. The neural network makes it easier. Actually, it would be closer to say that it makes it possible, because we can't wait around for a few thousand years while a supercomputer using conventional techniques churns out a predicted galactic catalogue for every iteration.

1 comment:

Unknown said...

Why do I get the feeling that if these guys applied their knowledge of ANN's to a different data set, they'd be arguing that they can make themselves very, very wealthy? Indeed, if ANN's were good at analyzing stochastic datasets, we'd have a lot of very wealthy AI students wandering about...

ANN's are a tool - but they aren't as accurate as they are reputed to be. I can't help but think their 80% to 90% accuracy might not get much better. Besides, ANN's get really complex, really fast; those guys dealt with a reasonably small system, and an obviously reasonably-sized dataset. (I did have some concerns about their ANN architecture, though. I don't think it could model unusual situations very well.)

Anyway - mix in some mapreduce with a distributed ANN (it would have to have a fairly complex architecture; I'd guess a multi-dimensional array of hidden nodes* would be needed) and it could shorten the time needed for modeling complex, stochastic, phenomena.

Neat article, though. After a hiatus of nigh on a decade, I've recently rekindled my interest in ANN's, so this was really interesting to read. Thanks for sharing! :-)

Carolyn Ann

* I once built such a beast. It was a long time ago, and the software I used just about melted with the complexity of the ANN. I remember the vendor going pale when I described what I'd tried...