On day zero of AISTATS, I gave a workshop on machine learning in astronomy, concentrating on the ideas of (a) trusting unreliable data and (b) the necessity of having a likelihood, or probability of the data given the model, making use of a good noise model. Before me, Zoubin Ghahramani gave a very valuable overview of Bayesian non-parametric methods. He emphasized something that was implied to me by Brendon Brewer's success on my MCMC High Society challenge and mentioned by Rob Fergus when we last talked about image modeling, but which has rarely been explored in astronomy: If you want to have the flexibility to discover correct structure in your data, you may have to adopt methods that permit variable model complexity. The issues are two-fold: For one, a sampler or an optimizer can easily get stuck in a bad local spot if it doesn't have the freedom to branch more model complexity somewhere else and then later delete the structure that is getting it stuck. For another, if you try to model an image that really does have five stars in it with a model containing only four stars, you are requiring that you will do a bad job! Bayesian non-parametrics is this kind of argument on speed, with all sorts of processes named after different kinds of restaurants. But just working with the simple dictionary of stars and galaxies, we could benefit from the sampling ideas at least.
In the evening Bernhard Schölkopf drove some people from his group and me up to the observatories at the summit on La Palma, where we did some observing with amateur equipment. His student Edgar Klenske is working on using a Gaussian Process to do far better real-time control of inexpensive (think unreliable) telescope tracking hardware. If they succeed, they could have a big impact not just for unreliable hardware but also reliable hardware.
No comments:
Post a Comment