Ramin Skibba at Inside Science wrote an article on deep learning and lens finding, and interviewed me about the work. Take a look here.
I’m searching for strong gravitational lenses. They are really interesting!
One of the tenets of the General Theory of Relativity is that space is curved by the presence of matter and energy. This curvature is what we experience as gravity – a fascinating subject in its own right. General relativity – a revolutionary and controversial theory in 1915 when Einstein published it – made the prediction that the curvature of spacetime caused by massive bodies like stars would deflect the passage of a ray of light*. Einstein did the calculations and predicted that a light ray passing very close to the sun would be deflected by a small amount – 1.3 arcseconds, or about 1/2800 of a degree. Thus, a star positioned just at the edge of the sun would appear to shift slightly to a different position in the sky because of this bending.
A concrete prediction like this one provides an opportunity to test a new theory in the court of empirical evidence. But the problem for Einstein was that stars are a little hard to see during the daytime – let alone when they are right next to the edge of the sun. Luckily, a total solar eclipse occurred in 1919 and this provided an opportunity to put this particular prediction to the test. The British astronomer Sir Arthur Stanley Eddington travelled to the island if Principe, where the total eclipse was visible, and took photographs of the sun from which the deflection could be measured. Of course, it was a match, and the prediction was borne out perfectly. It was Eddington’s expedition and subsequent vindication of the theory that grabbed headlines around the world and first made Einstein a household name. As an bonus, in a Europe exhausted by World War I, this collaboration of an English pacifist (Eddington was a Quaker) helping to prove a German theorist correct, was quite a showcase for the international society of science.
If the sun can bend the light of a star, what could a more massive body do? Through the 1920s and 1930s the existence of other galaxies and their place in the vastness of the cosmos came to be better understood. Among the astronomers who were thinking about the implications of general relativity, gravity and light was the Swiss Fritz Zwicky. One of Zwicky’s claims to fame is being the first to postulate the existence of dark matter, but in 1937 he suggested that galaxy clusters could act as gravitational lenses – bending distant light such that it was both distorted and magnified, just like an ordinary lens. The theory predicts multiple images of a distant light source might be present, and in some circumstances, the source could be stretched out into an arc or a complete circle – an “Einstein Ring”.
For a long time it was assumed by many that such an event would never be visible in practice; Einstein himself was of this opinion. What are the chances of things lining up so perfectly? Fortunately, thanks to telescope improvements and a growing understanding of the cosmos, lensing moved from the hypothetical to the possible in the minds of astronomers. However, it wasn’t until 1979 that a likely candidate was found, in the “twin quasar” Q0957+561. Walsh, Carswell and Weyman found what looked like two very similar objects close together in the sky, but due to the similarity of the objects, the fact that they happened to be at the same distance, and the presence of the possible lens galaxy in the middle, they suggested they might have found an elusive gravitational lens. Their discovery met with considerable skepticism but they were soon proved correct.
Today, we know of hundreds of lenses across the sky. However the universe is a big place, and those remaining to be discovered number in the tens of thousands or more as we see deeper and further back in time. Lenses are inherently beautiful and interesting, at least if you an astronomer, and so finding them and cataloguing them would be a worthwhile endeavour no matter what.
However, gravitational lenses turn out to be as useful to astronomers as their glass analogues in telescopes. Since the bending of light depends on the mass of the galaxy, if you can observe the bending, you know the mass – not just the gas and stars that make up the stuff we can see, but the dark matter that makes up the vast majority of the mass in the universe. Other aspects of lensed systems, such as the time delay between the different images of a multiply-images source (such as the “Twin QSOs” above), can help cosmologists probe the fundamental parameters of the universe such as the Hubble constant, H0.
Lenses are beautiful, interesting and scientifically useful, but rare. How do we find them from among hundreds of millions of galaxies on the sky? That’s my project, so stay tuned.
*It’s true there was a Newtonian prediction as well, based on the assumption that ‘corpuscles’ of light must have some mass. But the prediction deflection was half what GR predicted.
My project involves looking at images – lots of images. Not thousands, but hundreds of millions, so it’s lucky I have a computer to look through them for me. Getting it to do its job well is the goal of my research.
When astronomers work with image data they generally aren’t talking about jpegs or pngs. Professional astronomy left looking into eyepieces behind long ago and is now all about collecting data with the most sensitive instruments modern engineering can devise. When you think of an astronomical image you might imagine something like a giant wallpaper-quality picture of a galaxy, but a more scientifically cutting-edge galaxy might look more like this:
This is because a lot of the most interesting science is happening right on the edge of what we can detect, which in astronomy often means the furthest back in time that we can possibly see. At the edge, every bit of light you can measure is important, right down to the individual photon. In practice, an astronomical image taken from the instrument is usually distributed as a ‘FITS’ file which includes all sorts of metadata about where on the sky it’s looking, what filters are in place and the peculiarities of the instrument it was taken on. The data in the file is likely to be the actual count of electrons picked up by the detector, but if you are lucky it might be converted to flux values – the amount of energy detected per second per unit area. As you can imagine, the constraints are a little different when you are taking a picture of the sunset on an iPhone.
As you may know, when we look at an image on the screen and see colours we are seeing a mixture of red, green and blue channels of different brightnesses. Where the blue channel is bright and the others dim, we see blue, of course; but where they are all equally bright we see white or grey. Here again astronomical imaging is a little different. Astronomy is less concerned with what people can see and more with understanding the physical processes happening, so light is collected at all sorts of different wavelengths depending on the instrument and the nature of the investigation. This could mean radio waves, infra red, visible light, UV or even gamma rays to name a few. To make a colour image we can look at on a screen, you need red, green and blue channels; but which band (range of wavelengths) you use for each is up to you, and might not necessarily be what you think of as red, green and blue light if it includes, say, infra-red or ultra-violent information.
So, yesterday I was producing some colour images at random from the fields I was looking at while debugging a problem with a missing red band. Out of five different colour bands available, I was choosing three (g, r and i or y) to make colour images for testing, but some were coming out all blue (meaning the red channel was probably missing). Most of my little random images look like this – a smattering of fuzzy distant galaxies, many at the edge of what can be detected by this telescope:
Some have slightly more going on, like this patch – a bright elliptical galaxy in the centre, a handful of stars and at least a few spiral galaxies :
And then this guy wandered into my field of view.
Not only is it a beautiful sight, but that spiral galaxy helps put the other images in context. Almost everything we’re looking at there is a galaxy, they are just very far away. With this galaxy hovering in the foreground, it adds some perspective to those other images on the celestial sphere so far away in distant time and space.
The last couple of days I’ve been helping out – basically, as semi-skilled labour – on an exciting astronomy collaboration happening here in Melbourne and around the world. In astronomical terms, a million years is as the blink of an eye. When you are studying the expansion of the universe, the evolution of galaxies and the lifetimes of most stars, one is often talking in gigayears – time scales of billions of years.
However, there are astronomical phenomena that occur much more rapidly. Variable stars pulsate in hours and days; pulsars spin in seconds or less, ticking away over the aeons; supernovae explode in minutes and fade away in weeks and months. There may be other objects and phenomena of interest to science that operate on even shorter timescales. The problem is: If you aren’t sitting there, constantly watching the sky, how would you ever catch one?
The Deeper, Wider, Faster collaboration is designed to do just that and catch some of these fast-happening cosmic events. Astronomers in observatories around the world – Australia (including Jeff Cooke here at Swinburne and my fellow PhD students Igor Andreoni and Dany Vohl), the USA, Chile, and even Antarctica and in space – are all monitoring the same patch of sky, taking rapid exposures of a few minutes each of large fields using different parts of the electromagnetic spectrum. In order to catch transient (i.e. fast and temporary) events, a pipeline of software compares the new images with a template to see if anything has changed. By subtracting two images from one another, bits that remained the same should cancel out to nothing, but anything that has changed – whether brighter or dimmer – should stand out. These brighter or dimmer patches are then filtered to catch the most interesting ones, but ultimately need to be looked at by a human being to determine whether they are an artefact of the process (say, a bad pixel in the camera) or a real event somewhere in the universe.
The data the humans have to look at is complex in its simplicity. Ultimately, the object that will be found will be bright grey blobs in any single image. The problem is that artefacts caused by, say, a misaligned image subtraction or a ghost image in the electronics of the detector might also look like little blobs. Spotting the difference can be subtle and requires a bit of practice. Here’s a bad subtraction – some parts too bright, some too dim:
The middle of these two, on the other hand, just might be something:
Since the cameras are snapping pictures every few minutes, there’s more data always coming in to help verify candidates. In the case of a real transient, you’d expect the brightness to vary from image to image – either rising and falling or appearing and disappearing for the fastest of events. From this change in brightness over time, you can generate a light curve like the one pictured below for a closely-observed supernova:
Even with only 5 minutes between observations you can calculate how bright something is and make such a curve. Here’s one that popped up in real time in the control room:
Interesting – there is definitely something changing there!
Between the smooth blob picture and a rising or falling light curve a human can pretty quickly sort the wheat from the chaff, and a trained amateur can do after minutes or hours of training. So, what about learning machines? Although there is considerable sophistication here already in filtering out the junk so humans just look at the most promising candidates, there is clearly room for some deep learning to be applied here at the final stage.
There are two problems with throwing a neural network at this challenge. Firstly, a training set of good and bad candidates needs to be assembled so the computer has something to learn from. Secondly, one of the aims of Deeper, Wider, Faster is to find events that have never been observed before. So how could an algorithm learn to find things novel to science?
Like all things, the best answer is probably a mixture. For instance, I’m now sorely tempted to explore the issue of bad subtractions and see whether my artificial neural networks can get a handle on those. If that worked, then the humans would have more time to spend examining more promising candidates. At any rate, the project continues over the next week or so – let’s hope they find something novel and bizarre!
Welcome to “Deep Space, Deep Learning” where I will be blogging about my work at the intersection of machine learning and astronomy.
After a career in tech and politics I have been fortunate to have the opportunity to go back to my true love: science. Earlier this year I began my a PhD at the Centre for Astrophysics and Supercomputing at Swinburne University. CAS is a great centre with a buzzing vibe and lots of interesting work happening. I am fortunate to have two great supervisors in Prof Karl Glazebrook, an astronomer of some renown, and Dr Chris McCarthy, a computer vision researcher. What an interesting combo, something I will be exploring more here and in my work.
My area of research involves applying modern machine learning techniques to astronomy problems. In particular, I am training neural networks to search astronomy databases for gravitational lenses, one of the most amazing proofs of Einstein’s general theory of relativity.
For centuries, astronomy was a solitary affair. A single astronomer could painstakingly collect data over hours, days or years and then analyse it at her leisure. But the field is evolving rapidly. Modern astronomical instruments, and the surveys of the sky we conduct with them, are big science – they require ever larger amounts of money and ever-bigger collaborations. The volumes of data generated are entering the petabyte scale. How do we deal with so much interesting scientific data? Enter me and my little data science project!
On this blog I’ll be talking a bit about the Dark Energy Survey. DES is an amazing project, the collective work of hundreds of scientists looking to answer some fundamental questions about the nature of the cosmos. DES will map one eighth of the entire sky over 525 nights, each night collecting up to 2.5 terabytes of image data. This data, spanning hundreds of millions of galaxies, is rich in information about the history of the universe, but mining it is a data science challenge of the first order. I am fortunate to have been able to join DES and its team of hundreds of researchers around the world. I will be able to learn a lot from them, and if I’m lucky, contribute a little back to the project.
I’ll be back soon with more. In the next post I will look at how gravity bends light and how this can help us explore the history of the universe.