Today was the second day of Computing the Universe 2015 and Cosmo Hack Week. I spoke today about replacing frequentist correlation-function estimators with probabilistic inferences. I talked about likelihood-free inference (called "ABC" in previous posts, but I don't like that name), and about writing down a likelihood function based on Gaussian processes. Martin White (Berkeley) was my live critic, which was useful (and flattering).
There were many good talks; some highlights were the following: Cohn (Berkeley) talked about using unsupervised machine-learning methods to create features or descriptors of galaxy mass build-up histories, and then using a supervised regression to relate them to single-epoch observables; that seems like a productive way to think about designing new experiments. Budavari (JHU) spoke about hierarchical Bayes using GPUs to do the heavy lifting. He spoke mainly about GPU architecture and coding, but the hierarchical inference ideas he gave as his starting point are very relevant to Malz's projects; I have to get the two of them in touch. Font-Ribera (Berkeley) and Beutler (Berkeley) talked about BOSS large-scale structure results; the former about the IGM and baryon acoustic feature at redshift 2.5, and the latter about galaxies and the same feature at low redshift. Beautiful results in both cases. Schmittful (Berkeley) gave a beautifully clear talk on three-point functions and optimal estimators thereof. He has very nice results based on analysis of an Edgeworth expansion away from the Gaussian.
After the talks was hacking; at the hacking session I got together a team to argue out my ideas about CMB foregrounds. The discussion was very clarifying and I think there are some very simple baby steps towards a good project.
No comments:
Post a Comment