Source Separation with Deep Generative Priors

Authors: Vivek Jayaram, John Thickstun

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The method achieves state-of-the-art performance for MNIST digit separation. We introduce new methodology for evaluating separation quality on richer datasets, providing quantitative evaluation of separation results on CIFAR-10. We also provide qualitative results on LSUN. 5. Experiments We evaluate results of BASIS on 3 datasets: MNIST (Le Cun et al., 1998) CIFAR-10 (Krizhevsky, 2009) and LSUN (Yu et al., 2015).
Researcher Affiliation Academia * Equal contribution. Paul G. Allen School of Computer Science and Engineering, University of Washington. Correspondence to: Vivek Jayaram <vjayaram@cs.washington.edu>, John Thickstun <thickstn@cs.washington.edu>.
Pseudocode Yes Algorithm 1 BASIS Separation Input: m X, {σi}L i=1, δ, T Sample x1, . . . , xk Uniform(X) for i 1 to L do ηi δ σ2 i /σ2 L for t = 1 to T do Sample εt N(0, I) u(t) x(t) + ηi x log pσi(x(t)) + 2ηεt x(t+1) u(t) ηi σ2 i Diag(α) m g(x(t)) end for end for
Open Source Code Yes Code and instructions for reproducing these experiments is available online.1 1https://github.com/jthickstun/basis-separation
Open Datasets Yes We evaluate results of BASIS on 3 datasets: MNIST (Le Cun et al., 1998) CIFAR-10 (Krizhevsky, 2009) and LSUN (Yu et al., 2015).
Dataset Splits No The paper mentions evaluating results on specific numbers of separations (e.g., '6,000 pairs of equally mixed MNIST images', '25,000 separations (50,000 separated images) of two overlapping CIFAR-10 images') which are test evaluations. It does not explicitly state the train/validation/test data splits for the datasets used to train the models mentioned, nor for fine-tuning, beyond implicitly using test sets for evaluation.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments.
Software Dependencies No The paper mentions generative models (NCSN, Glow) but does not provide specific software dependencies with version numbers.
Experiment Setup Yes We adopt the hyper-parameters proposed by Song & Ermon (2019) for annealing σ2, the proportionality constant δ, and the iteration count T. The noise σ is geometrically annealed from σ1 = 1.0 to σL = 0.01 with L = 10. We set δ = 2 · 10−5, and T = 100. We find that the same proportionality constant between σ2 and η also works well for γ2 and η, allowing us to set γ2 = σ2.