Deep Learning for Classification of Astronomical Archives

点击翻译按钮获取中文摘要

Astronomy is known for producing highly appealing images which often get published and re-published world-wide. HST images, for instance, have appeared on web sites, news papers, and even stamps. Less well known is that these images are the tip of the ice berg. The large majority of images, even from HST, do not contain such pretty structures. Astronomers locate the news-worthy ones by manual inspection of all the data.Modern astronomical instruments provide far more data than can even be manually inspected. ALMA, for instance, can produce over 10,000 images from a single observation. Each of these measures a slightly different wavelength of radiation, and many of the images will be empty or near-empty. But without looking at each one, how does the astronomer know which images to select? This project will look at techniques to let computers do the work. They can investigate all images much faster than people can, and can learn which ones are of interest by looking for characteristic patterns. The problem is that they need to be taught what is 'characteristic' something even astronomers may find difficult to put into an algorithm. Malaysian scientists have unparalleled expertise in image classification using deep learning techniques, and in semantic descriptions. This expertise will be brought to the astronomical archives, to develop techniques of computer learning to aid astronomers. Manchester is one of the access points to the ALMA archive. The combination of the data and astrophysical expertise in Manchester and image classification by deep learning in Malaysia will provide new and powerful tools for science.