Celeste: Variational inference for a generative model of astronomical images
Authors: Jeffrey Regier, Andrew Miller, Jon McAuliffe, Ryan Adams, Matt Hoffman, Dustin Lang, David Schlegel, Mr Prabhat
ICML 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We check our approach on synthetic images. We also run it on images from a major sky survey, where it exceeds the performance of the current state-of-the-art method for locating celestial bodies and measuring their colors. The rest of the paper describes the Celeste model (Section 2) and its accompanying variational inference procedure (Section 3). Section 4 details our empirical studies on synthetic data as well as a sizable collection of astronomical images. |
| Researcher Affiliation | Collaboration | Jeffrey Regier, University of California, Berkeley JEFF@STAT.BERKELEY.EDU Andrew Miller, Harvard University ACM@SEAS.HARVARD.EDU Jon Mc Auliffe, University of California, Berkeley JON@STAT.BERKELEY.EDU Ryan Adams, Harvard University RPA@SEAS.HARVARD.EDU Matt Hoffman, Adobe Research MDHOFFMA@CS.PRINCETON.EDU Dustin Lang, Carnegie Mellon University DSTN@CMU.EDU David Schlegel, Lawrence Berkeley National Laboratory DJSCHLEGEL@LBL.GOV Prabhat, Lawrence Berkeley National Laboratory PRABHAT@LBL.GOV |
| Pseudocode | No | The paper describes the model and inference procedure mathematically but does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide any concrete access to source code for the methodology described. There are no links or explicit statements about code release. |
| Open Datasets | Yes | We also run it on images from a major sky survey, where it exceeds the performance of the current state-of-the-art method for locating celestial bodies and measuring their colors. Figure 1. An image from the Sloan Digital Sky Survey (SDSS, 2015) of a galaxy from the constellation Serpens, 100 million light years from Earth, along with several other galaxies and many stars from our own galaxy. For real astronomical images, ground truth is unknown. However, a region of the sky known as Stripe 82 has been imaged more than 30 times by modern telescopes |
| Dataset Splits | No | The paper does not provide specific details about training, validation, and test dataset splits (e.g., percentages, sample counts, or explicit standard split references) for the main model evaluation. It discusses |
| Hardware Specification | No | The paper mentions processing images of certain sizes and parallelization on separate processors but does not specify any exact hardware details such as CPU/GPU models, memory, or specific cluster specifications used for running its experiments. |
| Software Dependencies | No | The paper mentions using L-BFGS-B (Byrd et al., 1995) for optimization and refers to "Gaussian Mixtures.jl" (van Leeuwen, 2015) for fitting priors. However, it does not provide specific version numbers for these or other software dependencies, such as the programming language or other libraries, required for replication. |
| Experiment Setup | Yes | In principle, all parameters could be learned by variational inference. But we reuse some estimates from the existing photometric pipeline that are not thought to limit performance. The background noise level N bn is set by the existing photometric pipeline, based on a heuristic... The calibration constant N bn is set by first calibrating overlapping images relative to each other, and then by calibrating some images absolutely, based on benchmark stars (Padmanabhan et al., 2008). The image-specific parameters of the point spread function, . N nb; N nb; N nb/K k D1, are set by the existing photometric pipeline... Galaxy profiles parameters, .N ij ; N ij /J j D1, are set a priori too... To fit the prior on color .cs/ to existing catalogs, we use the expectation-maximization algorithm, initialized by k-means (van Leeuwen, 2015). Though D D 64 minimized held-out test error, we set D D 2, to work around a limitation of our present optimizer it only supports box constraints. We use L-BFGS-B (Byrd et al., 1995). When possible, we use existing star and galaxy catalogs for initialization. |