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].

Statistical Component Separation for Targeted Signal Recovery in Noisy Mixtures

Authors: Bruno Régaldo-Saint Blancard, Michael Eickenberg

TMLR 2024 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Then, we apply it in an image denoising context employing 1) wavelet-based descriptors, 2) Conv Net-based descriptors on astrophysics and Image Net data. In the case of 1), we show that our method better recovers the descriptors of the target data than a standard denoising method in most situations. Additionally, despite not constructed for this purpose, it performs surprisingly well in terms of peak signal-to-noise ratio on full signal reconstruction.
Researcher Affiliation Industry Bruno Régaldo-Saint Blancard EMAIL Center for Computational Mathematics Flatiron Institute New York, NY 10010 Michael Eickenberg EMAIL Center for Computational Mathematics Flatiron Institute New York, NY 10010
Pseudocode Yes Algorithm 1 Vanilla Statistical Component Separation Inputs: y, p(ϵ0), Q, T, gradient-based optimizer (e.g. LBFGS) Initialize: ˆx0 = y for i = 1 . . . T do sample ϵ1, . . . , ϵQ p(ϵ0) ˆL(ˆx0) = PQ k=1 ϕ(ˆx0 + ϵk) ϕ(y) 2 /Q ˆx0 one_step_optim h ˆx0, ˆL(ˆx0) i end for return ˆx0
Open Source Code Yes Codes and data are provided on Git Hub.2 2https://github.com/bregaldo/stat_comp_sep.
Open Datasets Yes We consider three different types of 256 256 images corresponding to a simulation of the emission of dust grains in the interstellar medium (the dust image), a simulation of the large-scale structure of the Universe (Villaescusa-Navarro et al., 2020) (the LSS image), and randomly selected images from the Image Net dataset (Deng et al., 2009) (the Image Net images).
Dataset Splits No The paper describes experiments involving different noise realizations of images but does not specify traditional train/test/validation dataset splits, as it focuses on denoising.
Hardware Specification Yes Each optimization takes 40 s with a GPU-accelerated code on a A100 GPU.
Software Dependencies No The paper mentions using a 'gradient-based optimizer (e.g. LBFGS)' and refers to the 'VGG-19_BN network' but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes We proceed similarly in the following, and fix the number of iterations to T = 30 and the batch size to Q = 100.5 ... We vary, for the colored noises, the amplitude σ of the noise considering 10 different levels ranging from 0.1 to 2.14 (logarithmically spaced) in unit of the standard deviation of x, and for the crosses noises, the density of crosses ρ considering 10 different values ranging from 0.001 to 0.063 (logarithmically spaced)6. ... with αi = 1/ P and P = 10σ ... and T = 10