Computers catching humans in galaxy classifications

An artificial neural net that mimics the network of neurons in the human brain has proven itself 90 percent as accurate as users of Galaxy Zoo when it comes to classifying galaxies, according to new research from scientists at the University of Cambridge and University College London. This new development will eventually allow Galaxy Zoo users to focus on the more interesting peculiar galaxies whilst computers characterise the billions of standard spiral and elliptical galaxies that will be uncovered by the next generation of surveys.

Computers are getting better are classifying galaxies. Image: NASA/ESA/Hubble Heritage Team (AURA/STScI).

An artificial neural net is an algorithm that takes inputs – in this case a galaxy’s shape, colour, compactness and texture – and draws relationships between them to ascertain what type of galaxy they are, the same way your eyes see four legs and a flat top and your brain works out that you are looking at a table. Artificial neural nets have been used in a wide variety of applications besides astronomy, from face-recognition software to stock market predictions.

The artificial neural net is ‘trained’ on a sample of galaxies that have already been correctly classified by human beings. Because each galaxy on Galaxy Zoo is classified by many users, the overall human result is considered accurate. The reason that the computer is only correct nine times out of ten is that peculiar galaxies – those that have no defined shape or which have properties atypical of their type – can throw the computer. “For example, we know that most elliptical galaxies are red in colour and most spirals are blue,” says Cambridge’s Dr Manda Banerji, who led the research that will be published in the Monthly Notices of the Royal Astronomical Society. The artificial neural net is ‘trained’ to recognise that red galaxies are likely elliptical, and blue galaxies probably spirals. However, Galaxy Zoo has turned up a small number of red spirals and blue ellipticals that still seem to be forming stars. “These kinds of objects may therefore not be correctly classified by the neural network,” Banerji told Astronomy Now.

Sixty million galaxy classifications have thus far been made on Galaxy Zoo by its 250,000 members in two years, but upcoming surveys such as the VISTA telescope’s Hemisphere Survey, and the Dark Energy Survey that will be conducted on the four-metre telescope at the Cerro Tololo Inter-American Observatory – both in Chile – hope to find far more galaxies than that. “We are now entering an age where galaxy surveys with hundreds of millions and even billions of galaxies will become routine,” says Banerji. This is far too many even for the dedicated folk at Galaxy Zoo, and with modern computing power the artificial neural net can classify a million galaxies in an hour. In future, Galaxy Zoo users will have more time to turn up more unusual and exciting objects such as Hanny’s Voorwerp, leaving the bog-standard classification to the computer.