Follow the crowd: Scientific research tapping into public intelligence

The sound somewhat reminded me of a drilling machine. Except that this sound had an emotional element. There were clicks, whistles, and grunts. Some sounded playful, others angry.

I was listening to pilot whales on Whale.fm (www.whale.fm), a website that invites people to help marine scientists decipher the language of pilot whales. The task is simple. There is a main sound clip and you have to find a match for it from 36 other options. A graphical representation of the sound — called spectrogram — is provided to help narrow down the choices.

Once you find a match, you move on to the next clip.

And what does the match indicate? According to experts, pilot whales are highly social animals, which group themselves in families. Their calls are distinct from other species of whales. Moreover, every family has its own ‘dialect’ which helps individual members identify their group. When users match two clips, it helps scientists classify the sound as part of either the larger pilot whale language or individual dialect.

Earlier, researchers would focus on a subset of the sound clips that they recorded, says Sander von Benda-Beckmann, who works with TNO, a Dutch research agency which has collaborated on Whale.fm. “It is also sometimes argued that the categorisation may be biased if a small number of people have looked at the data. So I suggested that we use crowdsourcing to address this problem.”

Within the first few weeks of the launch, 60,000 matches were made, mainly by people outside the scientific community. Even children were participating in the programme, says von Benda-Beckmann, adding that the project will run for six to nine months, and based on the results may be extended further.

Whale.fm is one of the many crowd science projects that tap into the collective intelligence of people to help scientists conduct their research. At Oxford University, researchers have started a website called Ancient Lives, asking people to transcribe symbols written on papyrus scrolls which are 2,000 years old. At Washington University geneticists came up with a game called Foldit, where people have to fold protein structures into three-dimensional shapes using their spatial reasoning. Galaxy Zoo presents users with images of distant galaxies and asks them to classify the galaxies according to their shape.

Crowd science projects tend to revolve around pattern recognition, which is one area where humans fare better than computers, says Oxford-based astrophysicist Chris Lintott, who started Galaxy Zoo. Referring to classification of galaxies he says, “A lot of effort has gone into developing computer programs, but they have about a 70% success rate. That sounds good, but the problem is that there’s a bias — for example, the computer is more likely to get dusty spirals wrong, and that has a profound impact on our results. The truth is that humans are extremely good at pattern recognition and it’s hard to beat them. Think about how difficult it’s been to develop decent face-recognition software, while you or I can walk into a crowded room and instantly pick out our friends.”

The images used in Galaxy Zoo illustrate Lintott’s argument. The pictures, which come from the Hubble telescope, are often blurry. It needs some (human) imagination to detect, say, the spiral arms of a galaxy. Also, a computer algorithm trained to identify shapes may well miss other anomalies in the picture, which a human won’t.

In 2008, Hanny Van Arkel, a Dutch schoolteacher, posted an image on the Galaxy Zoo forums pointing out a strange green object near a galaxy. It was the first time astronomers had seen anything like that. The object was named Hanny’s Voonwerp, and scientists are now studying what this object is, and if there are other voonwerps in space.

Results thrown up by Galaxy Zoo are now being used by astrophysicists in a wide range of projects. For instance, scientists have discovered that round green galaxies called peas, which have turned out to be the most efficient stellar factories in the universe.

Last year, scientists at Washington University announced that one team playing Foldit had deciphered the folding of a protein crucial to research on AIDS.

Meanwhile, Devin Bloom, a PhD candidate in evolutionary biology at the University of Toronto Scarborough, uploaded pictures of more than 5,000 species of fish collected from the remote jungles of Guyana on Facebook. Bloom’s permission to stay in the country was about to expire and he needed the results in a few days. To Bloom’s surprise, within 24 hours all fishes got identified. “Social networking is so powerful, and scientists should be using it more to connect with the world at large,” he later told Science Daily.
Crowdsourcing will also help improve computer algorithms. “I think an ideal outcome would be for machines to take over - in fact, we’ve shown that the kind of data produced by citizen science projects is particularly useful in improving machine learning - but I don’t think that’ll happen soon. Our next step is to build a system that uses both machine and human classification, allowing the machines to call in human help if they need extra training or if something unusual is found.”