Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Variational Inference for Deblending Crowded Starfields
Authors: Runjing Liu, Jon D. McAuliffe, Jeffrey Regier, The LSST Dark Energy Science Collaboration
JMLR 2023 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In our experiments with SDSS images of the M2 globular cluster, Star Net is substantially more accurate than two competing methods: Probabilistic Cataloging (PCAT), a method that uses MCMC for inference, and DAOPHOT, a software pipeline employed by SDSS for deblending. |
| Researcher Affiliation | Collaboration | Runjing Liu EMAIL Department of Statistics University of California, Berkeley Berkeley, CA 94720, USA Jon D. Mc Auliffe EMAIL The Voleon Group Berkeley, CA 94704, USA and Department of Statistics University of California, Berkeley Berkeley, CA 94720, USA Jeffrey Regier EMAIL Department of Statistics University of Michigan Ann Arbor, MI 48109, USA |
| Pseudocode | No | The paper describes the generative model, variational distribution, and inference procedure in text and with equations, but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code to reproduce our results is publicly available in a Git Hub repository (BLISS, 2023). |
| Open Datasets | Yes | We use images from the Sloan Digital Sky Survey (SDSS) of the globular cluster M2... The M2 globular cluster was also imaged in the ACS Globular Cluster Survey (Sarajedini et al., 2007) using the Hubble Space Telescope (HST)... We next demostrate Star Net on a larger region of the sky. The DECam survey imaged stars in our own Milky Way... |
| Dataset Splits | No | The paper describes training Star Net on images sampled from a generative model and then evaluating it on specific astronomical images (e.g., M2 100x100 subimage, DECam 4000x2000 frame). It does not specify training, validation, or test splits for a fixed dataset, but rather generates data on-the-fly for training and uses distinct real-world images for evaluation. |
| Hardware Specification | Yes | On a single NVIDIA GeForce RTX 2080 Ti GPU, this fitting procedure took one hour. |
| Software Dependencies | No | The paper mentions 'Adam' as an optimizer and 'Optuna' for hyperparameter optimization but does not provide specific version numbers for these or any other software libraries or programming languages. |
| Experiment Setup | Yes | We ran SGD to minimize the expected forward KL for 400 epochs; at each epoch, 200 images of size 100x100 pixels were sampled from the generative model. We performed optimization with Adam (Kingma and Ba, 2014)... The first convolutional layer (green block, Figure 3) has 17 output-channels, a kernel size of three, a stride of one, and one pixel of padding. All convolutional layers inside residual block 1, as well as the convolutional layers on the top row of residual block 2 (Figure A.7) also have the same parameters... Inside the residual blocks, the dropout layers have dropout probability of 0.11399. The final fully connected block (red block, Figure 3) has latent dimension 185, and a dropout probability of 0.013123. |