The AstroStat Slog

I’ve been complaining about how one can do machine learning on solar images without a training set? (see my comment at the big picture). On the other hand, I’m also aware of challenges in astronomy that data (images) cannot be transformed freely and be fed into standard machine learning algorithms. Tailoring data pipelining, cleaning, and processing to currently existing vision algorithms may not be achievable. The hope of automatizing the detection/identification procedure of interesting features (e.g. flares and loops) and forecasting events on the surface of the Sun is only a dream. Even though the level of image data stream is that of tsunami, we might have to depend on human eyes to comb out interesting features on the Sun until the new paradigm of automatized feature identification algorithms based on a single image i.e. without a training set. The good news is that human eyes have done a superb job!

From A Survey of the Statistical Theory of Shape by David G. Kendall, Statistical Science, Vol. 4, No. 2 (May, 1989), pp. 87-99.
It is well known that no classical test for two dimensional stochastic point processes can match the performance of the human eye and brain in detecting the presence of improbably large holes in the realized pattern of points. This fact has generated a great deal of research in the last few years, especially in connection with the large “voids” and long “strings” that the eye sees (or declares that it sees) in maps of the Shane and Wirtanen catalogue of positions of galaxies. Astronomers are interested in (i)whether these phenomena are sufficiently extreme to require explanation, and if so (ii) whether any of the various “model” universes now in vsgue can be said to display them to just the same degree.