The Crowd and the Cosmos: Adventures in the Zooniverse

by Lintott, Chris

  that are relatively mainstream—some tweak of the vacuum

  energy, a scalar field left over from the start of the Universe—to

  the speculative. The latter possibilities are maybe the most excit-

  ing, often involving a direct challenge to the foundations of

  Einstein’s general relativity and its ideas about how gravity

  works. (Not a game that traditionally ends well for the challenger,

  but I suppose there’s always some hope that this time we’ll catch

  Albert out.)

  To concentrate, as a thousand articles and not a few books

  have done, on the intellectual games of theorists, no matter how

  diverting or elegant they are, is to miss the point.

  One of these ideas will win out, but not because of a sudden

  theoretical breakthrough. There will be no dropped chalk at the

  end of a lecture to a stunned audience, and no one will be leaping

  out of a bath shouting ‘Eureka!’ What’s needed, desperately, is

  more data. If we knew, for example, that the strength of dark

  energy was changing over time, that would rule out many of the

  available theories and give researchers something to aim for.

  What’s more, the observational route to this is clear. What we need

  are more supernovae. The discovery of more distant examples

  would mean that we could compare the past effect of dark energy

  with present-day values, and the discovery of more nearby explo-

  sions means that we can take better account of systematic effects.

  This search has recently become even more important. Just as

  it looked like the results from many different cosmological

  162 From Supernovae to Zorill aS

  probes were converging on a single solution to the parameters

  that control the Universe’s evolution, an intriguing set of results

  suggest we might not be done yet. Measures of Hubble’s con-

  stant- the rate of expansion of the present day Universe made

  with supernovae are giving a higher answer to those derived

  using methods which depend on measurements of the cosmic

  microwave background, which suggest the Universe is expand-

  ing more slowly. In a well-behaved Universe, the two should

  agree with each other, so this is puzzling.

  It could be that there’s nothing to worry about. The difference

  is small enough that the ‘tension’, as it’s coyly termed, could just

  be due to chance, similar to flipping a coin and getting three

  heads in a row. In such circumstances, it’s probably premature to

  conclude that the coin is biased towards heads. In the case of

  these separate measurements, there seems to be a little more

  than a 1 in 100,000 chance of the difference being a coincidence;

  not enough for scientists to be sure that’s it’s real, but certainly

  enough to worry about. Whole conferences have now been

  dedicated to the problem, and both groups—those that study

  supernovae and those who stare at the cosmic microwave back-

  ground—are adamant that there’s no simple explanation. Either

  we don’t understand the early Universe properly, or something is

  seriously wrong with our cosmological models, or supernovae

  are odder than we think.

  All of those possibilities are exciting, and in each case we need

  more data, and so both understanding this intriguing result and

  getting a critical clue to the nature of dark energy—the key prob-

  lem in twenty-first-century physics— depends on our ability to

  find changes in the sky. We’ve already seen that finding planets—

  and maybe (though probably not) aliens—is essentially a prob-

  lem of watching things change. So is keeping the Earth safe from

  killer asteroids, or detecting the relics of the earliest days of the

  From Supernovae to Zorill aS 163

  Solar System that lurk beyond Pluto. Our understanding of stars

  (and whether they can support planets with intelligent life)

  depends on understanding their variability; the Sun has sun-

  spots, and some other stars at least have starspots. And it’s not

  too much to hope that one day soon we might watch the centres

  of nearby galaxies flicker as material falls into their central black holes.

  Given this long to-do list, it’s not surprising that telescopes all

  over the world are being converted to look for new transients.

  There is hardly a modern survey that doesn’t have looking for

  changes in the sky as at least one of its goals, but to do the job

  properly dedicated facilities are needed. Optical systems need to

  be stable to allow images taken days, months, or even years apart

  to be compared, and instruments on the look-out for changes

  need to have as wide a field of view as possible. In the most

  extreme cases, cameras and computers might need to work fast

  to trigger alerts so that other telescopes can follow up on

  discoveries.

  To get an idea of what a modern transient-hunting machine

  might look like, you could travel to a hitherto obscure peak in

  the Atacama Desert in northern Chile. The desert has long been

  recognized as one of the best places on Earth for observatories,

  high above the often cloudy coast but lower than the snowy

  peaks of the Andes. Look at a satellite photo of the area, and the

  most typical sight is a clear strip between two belts of cloud, and

  it’s here that many of the world’s largest telescopes are placed.

  On a mountain called Cerro Pachón, construction of a new eye

  on the sky is underway.

  This is the Large Synoptic Survey Telescope, or LSST, whose

  mirror I encountered earlier in the University of Arizona’s sur-

  real football stadium-based mirror lab. Staring at a distorted ver-

  sion of myself in the newly shiny surface, it was hard to imagine

  164 From Supernovae to Zorill aS

  that it would ever get anywhere near being ready to ship data to

  the world’s astronomers, but now first light—the moment when

  the first pictures of the sky are taken—is just around the corner.

  The initial images with a temporary, commissioning camera are

  due in 2021, and the survey proper will start, all being well, in

  2023.

  It’s hard for me not to be slightly scared by the prospect. The

  roots of the LSST project go back almost two decades, when the

  first plans for such a telescope were hatched. Even then, it was

  clear that for all the clever optics, the biggest challenge would be

  dealing with the data such a survey would produce. After full

  operations start, LSST should produce about thirty terabytes of

  images a night, more each night than the Hubble Space Telescope

  produced in its first fifteen years. That’s just the static images,

  though, and it’s the numbers of expected transients which are

  which are truly frightening.

  Nothing like LSST has ever been built before, so predictions

  are uncertain, but subscribing to a service that provided a text

  message every time LSST detects a change would leave you

  waking up to at the very least a million text messages every

  clear night. Most would be routine changes—viewed with a

  telescope as large as LSST, a very large number of
stars will vary

  in brightness—but hidden in the stream will be everything you

  can imagine. If type 1a supernovae are your thing, there will be

  plenty hidden in the data if you can only find them.

  One solution is to depend on machine learning. Scientists all

  over the world are preparing ‘brokers’, little software helpers

  which will listen to the great stream of data flowing from the

  observatory and shout loudly when they spot something inter-

  esting. For the most interesting or useful transients, I suspect

  we’ll see competing brokers from different teams, announcing

  the highlights from LSST’s transients, either loudly to their world

  From Supernovae to Zorill aS 165

  or more quietly to their creators. Like an electronic version of an

  old trading hall, the advantage will accrue to those who can best

  filter information or make sense of the cacophony.

  Not all those acting as brokers will be machines. Where there

  is a wealth of data to be organized and sorted, the experiences

  recounted earlier in this book have taught me that there might be

  a place for citizen science; the Zooniverse hosted its first super-

  nova-hunting project back in 2010. The data rate, provided by a

  reconditioned telescope on Palomar Mountain in California now

  pressed into service as the Palomar Transient Factory, was a little

  more tractable, but the principle was the same. The telescope

  scanned the sky, and a computer checked each night’s images

  against a set of standard images. The few thousand such candi-

  dates a night were uploaded to our website, and a dedicated band

  of a few thousand volunteers jumped onto a dedicated website

  each day to sort through them.

  They were fast, collectively analysing a night’s worth of data in

  just fifteen minutes, and they were accurate. We were able to

  broadcast their classifications to observers stationed around the

  world, and add newly confirmed supernovae to the cosmological

  harvest. As the survey progressed, though, these classified super-

  novae also provided new training data for use by would-be

  supernova-hunting robots. Eventually, an extremely bright stu-

  dent in Berkeley, California produced a trained machine-learning

  solution that performed accurately enough to satisfy the astron-

  omers running the survey, and as they preferred clinical algorith-

  mic precision to messy and confounding citizen science our

  project was no more.

  I’ll return to this project later, as I think the experience of the

  volunteers who took part has much to tell us about the future of

  citizen science in general. For now, though, let’s continue to

  think like transient-hunting scientists, and worry about getting

  166 From Supernovae to Zorill aS

  hold of as much data as possible. Proof that relying solely on

  their machine was a mistake arrived at Earth on 21 January 2014,

  and was first announced by an unusual team from a truly unlikely

  place.

  London is a terrible spot to put an observatory. If you had to

  pick a site among the glitz and glare of the brightly lit metropolis, the very worst place would be in the centre of the West End. The

  second worst, though, would be along one of the capital’s main

  roads—alongside the A1 as it cuts through built-up North London,

  for example. Yet if you drive north on the A1 and look left just at

  the right time somewhere in Edgware, you’ll spot the gleaming

  domes of the University of London’s Mill Hill observatory.

  It’s a long way from a pristine Chilean mountain top, but that’s

  OK. The observatory exists primarily as a teaching tool, giving

  students on astrophysics courses at University College London

  experience in carrying out astronomical observation and data

  reduction. While the largest telescope is still the beautiful

  Radcliffe refractor, now more than a century old, it’s the modern

  telescopes clustered around it that get the most use.

  Back on that fateful January night, Steve Fossey—doyen of the

  observatory’s teaching labs since well before I was a PhD student

  at University College London—was scheduled to give a practical

  introduction to the telescopes to a bunch of undergraduates.

  Light pollution isn’t the only problem with the site, though, and

  clouds closed in overhead as the session was getting going. As the

  students took a break with pizza, Steve slewed one of the smaller

  telescopes over to one of the last clear patches, a region in Ursa

  Major that contains the nearby galaxy M82.

  M82 is known as the cigar galaxy—it is a spiral viewed almost

  edge on, presenting itself as a thin needle of light on the sky. As

  that night’s image appeared on the screen, Steve noticed a new,

  bright star located at one end of the disc, something that definitely

  From Supernovae to Zorill aS 167

  wasn’t in archived images of the same galaxy. From eating take-

  away pizza, four students—Ben Cooke, Guy Pollack, Tom

  Wright, and Matt Wilde—were suddenly following up on what

  proved to be the supernova discovery of the twenty-first century

  so far. The clouds were closing in, and the students and Steve

  rushed to get confirmation images using filters that exposed the

  camera to different colours. Clinching evidence came when they

  used a second telescope on the site to take an image of the same

  galaxy, and saw that the supernova was still there. It wasn’t an

  instrumental error, or something weird happening in the cam-

  era; it was real (Plate 9).

  From there things moved fast. The standard procedure is to

  report such a discovery to the wonderfully named Central Bureau

  for Astronomical Telegrams in the US, who announced the dis-

  covery to the world. Within hours of the initial discovery, tele-

  scopes around the world had observed what the University

  College London team had found, confirming that supernova

  2014J (as it was now known) was not only real but a type 1a. The

  opportunity to study this most important type of explosion up

  close—or at least at a distance of only eleven and a half million

  light years—was unprecedented, and it became one of the most

  observed objects of the twenty-first century.

  The discovery of such an object by pizza-munching students

  is a great story, and it was wonderful to see Steve’s sharp eyes get

  some recognition, but the truth is that they should never have

  had a chance. Automated surveys had caught the supernova

  before it was observed in London, but the routines used to scan

  for interesting transients didn’t catch it and so didn’t sound the

  alarm. This seems odd. The supernova is incredibly obvious in

  images—it’s the bright star that wasn’t there in 2013—but this is

  only true for human observers. My guess is that the training sets

  used to send machines hunting for transients didn’t include

  168 From Supernovae to Zorill aS

  anything this bright, and so the computers had ‘learned’ that

  anything that obvious couldn’t possibly be real. And so the

  supernova rema
ined unfound.

  There are, to my surprise, at least ten images of the M82 super-

  nova from before the discovery, including some from amateur

  astrophotographers who either didn’t process their data straight

  away or who didn’t know the galaxy well enough to recognize

  that the star was new. Mostly it was the former; even amateur

  astronomers with (advanced) backyard telescopes now leave

  looking at the images to the daytime rather than viewing them as

  they come in. Such images are, after the fact, still incredibly use-

  ful; it turned out that the rise to peak brightness was more rapid

  for this event than normal, indicating some unexpected process

  at play, a result that is still causing debate among experts.

  One of the surveys that imaged M82 during the period when

  the supernova was visible but not yet known was the Palomar

  Transient Factory which fed data to the Zooniverse’s supernova

  project. Had our supernova project still been operating I’m sure

  we would have caught it, and quickly. There is, of course, no real

  impediment to adapting the machine-learning routines used to

  include objects like this one; if I’m right that it was the bright-

  ness that made it difficult, one could simply train on as many

  bright supernovae as necessary. (If there are enough examples

  in both the Universe and our survey of it, that is. I’ll talk about

  truly rare objects a little later.) The point, though, isn’t that one couldn’t possibly have designed a system which wouldn’t have

  failed in this way, it’s that no one did. When dealing with a com-

  plex problem, and real, messy, noisy data, anticipating every

  possible eventuality is difficult. Ensuring that every case is

  covered, every loophole closed, and every unusual object antici-

  pated is impossible. Preparing a training set that reflects reality

  is next to impossible.

  From Supernovae to Zorill aS 169

  We could continue to bet on improvement in machine learn-

  ing. We might have missed this supernova, for example, but there

  will be others. Certainly there’s plenty of research funding going

  into making such systems work better. I prefer, though, to

  acknowledge the limits of any system we build and look to com-

  bine the best of automated scanning, which brings speed and

  consistency, and the quirky responses of adaptable citizen scien-

  tists, capable of going beyond their training.