Sampling Binary Data by Denoising through Score Functions
Authors: Francis Bach, Saeed Saremi
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate our formalism and theoretical findings by experiments on synthetic data and binarized images. 5. Experiments We now study how our new sampling scheme operates, first on synthetic data to understand the role of the noise parameter α and step-size η, then on binarized MNIST digits. 5.2. Binarized MNIST In this section we present our experiments on MNIST (Le Cun et al., 1998). |
| Researcher Affiliation | Collaboration | Francis Bach 1 Saeed Saremi 2 1Inria, Ecole Normale Sup erieure, PSL Research, University, Paris, France 2Frontier Research, Prescient Design, Genentech. Correspondence to: Francis Bach <EMAIL>, Saeed Saremi <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 The single-measurement one-stage discrete Langevin algorithm. Algorithm 2 The single-measurement two-stage discrete Langevin algorithm. Algorithm 3 Multi-measurement binary sampling via single-stage discrete Langevin algorithm (Alg. 1) in the inner loop. |
| Open Source Code | No | The paper does not contain any explicit statement about the release of source code or a link to a code repository for the methodology described. |
| Open Datasets | Yes | We also conduct experiments on binarized MNIST by qualitatively studying the role of α... Binarized MNIST In this section we present our experiments on MNIST (Le Cun et al., 1998). |
| Dataset Splits | No | The paper mentions using a 'test set' for MNIST but does not provide specific details on the training/validation/test split percentages, sample counts, or a citation to the specific split methodology used. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used to run the experiments, such as GPU/CPU models or other computing infrastructure specifications. |
| Software Dependencies | No | The paper mentions using 'Adam W' for optimization but does not provide specific version numbers for any key software components, libraries, or programming languages used in the implementation. |
| Experiment Setup | Yes | For optimization, we used Adam W (Loshchilov & Hutter, 2019) with the constant learning rate of 10 4 and the weight decay of 10 2. The step-size η is set to 1/α in all experiments. We present our experiments for α {0.25, 0.5, 1, 2} in Figs. 5 and 6. The denoising is set up using logistic regression as outlined in Section 2.3, where we parametrize fθ using the U-Net architecture (Ronneberger et al., 2015) with the modifications by Dhariwal & Nichol (2021). |