KIPAC astrophysicists have used a technique that processes information in a way analogous to the human brain in order to determine whether galaxy shapes can help determine their place in the Universe.
The effect of adding multiple parameters representing galaxy shape information on the photo-z accuracy, as determined by Singal et al. with their neural network method.
One of the major technical challenges for cosmology in the coming years will be the problem of photometric redshift. Photometric redshift estimation, or 'photo-z', is the process of discerning the redshift - and thus the relative distance - of a galaxy from the measured flux in a few distinct color bands, rather than in the traditional way with a full optical spectrum of the galaxy where spectral lines are clearly visible. The upcoming era of optical large cosmological surveys, including the Large Synoptic Survey Telescope (LSST), will rely on photo-z estimation in using data to determine parameters such as the density and evolution of dark energy.
The determination of cosmological parameters and other scientific goals of large surveys are particularly sensitive to errors in photo-z estimation, leading to the necessity of reducing the number of 'outliers' - those galaxies whose estimated photo-z is far from the actual redshift. The astronomical community has been anticipating the challenge for several years with many groups pursuing strategies to improve photo-z estimation. One such strategy that has been suggested is the inclusion of information relating to the apparent shape of galaxies in addition to the band fluxes. This is a reasonable hypothesis as it is accepted that galaxies change shape over time, and higher redshift galaxies are more distant and so we are seeing them as they were farther back in time.
KIPAC scientists Jack Singal, Marina Shmakova, and Brian Gerke have carried out a determination of whether inclusion of shape parameters of galaxies can improve the photo-z accuracy. The team used a technique called an artificial neural network to determine the photo-zs, which is ideal for elucidating whether there is additional useful information in the galaxy shapes. Artificial neural networks are information processing computer programs that are inspired by the way the human brain recognizes patterns. The network consists of mathematical 'neurons' and formulas of 'weights' that connect them, allowing one group of neurons to pass a signal to another, with the output value in the final neuron corresponding to the redshift. The network is 'trained' on a set of galaxies with input parameters and specified known redshift, and the internal connecting weights are adjusted iteratively so that the network as reliably as possible associates any given input parameters with the correct redshift. Once trained and with the weights set, then the network can be used to estimate the redshifts of other galaxies not in the training set, given the same type of inputs. If the process is repeated with and without the inclusion of galaxy shape information, the usefulness of this information for photo-z estimation can be determined.
The KIPAC team, using an artificial network designed and programmed by Singal, found that, surprisingly, the inclusion of galaxy shape information does not significantly improve the photo-z accuracy, and in fact too much shape information can be detrimental. It seems that any true additional relevant information related to redshift contained in the galaxy shape information is exceeded by the amount of redundant information. The finding is consistent with a few previous determinations by others, but the novel aspect is the use of the neural network technique and a sample of galaxies spanning a larger range of redshifts, and therefore a more comprehensive understanding of the potential for useful information in galaxy shapes. This result may help guide the astronomical community in strategies to approach the photo-z problem.
This work is based on a paper submitted to the Publications of the Astronomical Society of the Pacific and available from astro-ph at arXiv:1101.4011.