One of the joys of working in the Zooniverse is the sheer variety of people who are interested in our work, and I spent a happy couple of days toward the end of last year at a symposium about Discovery Infomatics – alongside a bunch of AI researchers and their friends who are trying to automate the process of doing science. I don’t think they’d mind me saying that we’re a long, long way from achieving that, but it was a good chance to muse on some of the connections between the work done by volunteers here and by our colleagues who think about machine learning.
In the past we’ve shown that machines can learn from us, but we’ve also talked about the need for a system that can combine the best of human and machine.
I’m still convinced that that will especially be needed as the size of datasets produced by scientific surveys continues to increase at a frightening pace. The essential idea is that only the proportion of the data which really needs human attention need be passed to human classifiers; an idea that starts off as a non-brainer (wouldn’t it be nice if we could decide in advance which proportion of Galaxy Zoo systems are too faint or fuzzy for sensible decisions to be made?) and then becomes interestingly complex.
This is particularly true when you start thinking of volunteers not as a crowd, but as a set of individuals. We know from looking at the data from past projects that people’s talents are varied – the people who are good at identifying spiral arms, for example, may not be the same people who can spot the faintest signs of a merger. So if we want to be most efficient, what we should be aiming for is passing each and every person the image that they’d be best at classifying.
That in turn is easy to say, but difficult to deliver in practice. Since the days of the original Galaxy Zoo we’ve tended to shun anything that resembles a test before a volunteer is allowed to get going, and in any case a test which thoroughly examined someone’s ability in every aspect of the task (how do they do on bright galaxies? on faint ones? on distant spirals? on nearby ellipticals? on blue galaxies? what about mergers?) wouldn’t be much fun.
One solution is to use the information we already have; after all, every time someone provides a classification we learn something not only about the thing they’re classifying but also about them. This isn’t a new idea – in astronomy, I think it’s essentially the same as the personal equation used by stellar observers to combine results from different people – but things have got more sophisticated recently.
As I’ve mentioned before, a team from the robotics group in the department of engineering here in Oxford took a look at the classifications supplied by volunteers in the Galaxy Zoo: Supernova project and showed that by classifying the classifiers we could make better classifications. During the Discovery Infomatics conference I had a quick conversation with Tamsyn Waterhouse, a researcher from Google interested in similar problems, and I was able to share results from Galaxy Zoo 2 with her*.
We didn’t get time for a long chat, but I was delighted to hear that work on Galaxy Zoo had made it into a paper Tamsyn presented at a different conference. (You can read her paper here, or in Google’s open access repository here.) Her work, which is much wider than our project, develops a method which considers the value of each classification based (roughly) on the amount of information it provides, and then tries to seek the shortest route to a decision. And it works – she’s able to show that by applying these principles we would have been done with Galaxy Zoo 2 faster than we were – in other words, we wasted some people’s time by not being as efficient as we could be.
That doesn’t sound good – not wasting people’s time is one of the fundamental promises we make here at the Zooniverse (it’s why we spend a lot of time selecting projects that genuinely need human classifications). Zoo 2 was a long time in the past, but knowing what we know now should we be implementing a suitable algorithm for all projects from here on in?
Probably not. There are some fun technical problems to solve before we could do that anyway, but even if we could, I don’t think we should. The current state of the art of such work misses, I think, a couple of important factors which distinguish citizen science projects from other examples considered in Tamsyn’s paper particularly. To state the obvious: volunteer classifiers are different from machines. They get bored. They get inspired. And they make a conscious or an unconscious decision to stay for another classification or to go back to the rest of the internet.
The interest a volunteer will have in a project will change as they move (or are moved by the software) from image to image and from task to task, and in a complicated way. Imagine getting a galaxy that’s difficult to classify; on a good day you might be inspired by the challenge and motivated to keep going, on a bad one you might just be annoyed and more likely to leave. We all learn as we go, too, and so our responses to particular images change over time. The challenge is to incorporate these factors into whatever algorithm we’re applying so that we can maximise not only efficiency, but interest. We might want to show the bright, beautiful galaxies to everyone, for example. Or start simple with easy examples and then expand the range of galaxies that are seen to make the task more difficult. Or allow people a choice about what they see next. Or a million different things.
Whatever we do, I’m convinced we will need to do something; datasets are getting larger and we’re already encountering projects where the idea of getting through all the data in our present form is a distant dream. Over the next few years, we’ll be developing the Zooniverse infrastructure to make this sort of experimentation easier, looking at theory with the help of researchers like Tamsyn to see what happens when you make the algorithms more complicated, and talking to our volunteers to find out what they want from these more complicated projects – all in our twin causes of doing as much science as possible, while providing a little inspiration along the way.
* – Just to be clear, in both cases all these researchers got was a table of classifications without any way of identifying individual volunteers except by a number.