In plain words: A deep learning program reads 3D dark matter maps to predict key numbers about our universe, built to train across supercomputer nodes at once. Unlike single-machine training, it spread the work over 8,192 nodes with little slowdown and predicted those numbers more accurately than before.
Abstract
Deep learning is a promising tool to determine the physical model that describes our universe. To handle the considerable computational cost of this problem, we present CosmoFlow: a highly scalable deep learning application built on top of the TensorFlow framework. CosmoFlow uses efficient implementations of 3D convolution and pooling primitives, together with improvements in threading for many element-wise operations, to improve training performance on Intel(C) Xeon Phi(TM) processors. We also utilize the Cray PE Machine Learning Plugin for efficient scaling to multiple nodes. We demonstrate fully synchronous data-parallel training on 8192 nodes of Cori with 77% parallel efficiency, achieving 3.5 Pflop/s sustained performance. To our knowledge, this is the first large-scale science application of the TensorFlow framework at supercomputer scale with fully-synchronous training. These enhancements enable us to process large 3D dark matter distribution and predict the cosmological parameters $Ω_M$, $σ_8$ and n$_s$ with unprecedented accuracy.
Amrita Mathuriya, Deborah Bard, Peter Mendygral, Lawrence Meadows, James Arnemann, Lei Shao, Siyu He, Tuomas Karna, Daina Moise, Simon J. Pennycook, Kristyn Maschoff, Jason Sewall, et al.
arXiv:1808.04728 · astro-ph.CO, astro-ph.IM, cs.LG, physics.comp-ph · submitted Aug 14, 2018 · updated Nov 9, 2018
abstract · pdf · html · 11 pages, 6 pages, presented at SuperComputing 2018
By going to the trouble of making a principled forward model including error statistics, you get error bars on all reported parameters, such as the age of the universe - these error bars are critical to later analysis. ("This model requires a Hubble constant in the range ... which is not supported by current observational evidence.")
Parameter estimation from large scale structure is not as "easy". OTOH, I wonder about the accuracy of their map from (cosmological parameters) -> (Large scale structure), because that's what their system is inverting. That is, I wonder about their training data.
From sec. 4C of the current paper, and comparing to their reference to Ravanbakhsh et al 2017 (ICML, [1]), they seem to be using numerical simulations (N-body sims representing gravitational coalescence). These models might have large systematic errors, and their inverse may have its own problems.
Long way to say: I wonder how these estimates of Lambda CDM parameters are viewed by the astrophysics community? The present paper is more of a systems/methods paper than an astrophysics paper. The estimates will come with no error bars, and will be critically based on the simulations used, and the NN limitations.
[1] https://arxiv.org/pdf/1711.02033.pdf