A noise-corrected Langevin algorithm and sampling by half-denoising
Authors: Aapo Hyvarinen
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4 Experiments Data models We use two different scenarios. The first is Gaussian mixture model in two dimensions. ... MCMC algorithms The starting point is that we only know the noisy-data score function (i.e. the score function for noisy data), and for one single noise level. ... Results 1: GMM bias removal Basic visualization of the results is given in Figs 1 and 2, which we will not comment any further. A comprehensive quantitative comparison is given in Fig. 3. |
| Researcher Affiliation | Academia | Aapo Hyvärinen EMAIL Department of Computer Science University of Helsinki, Finland |
| Pseudocode | No | The paper describes the proposed algorithm using mathematical equations (Eq. 6, 7, 15, 16) and textual descriptions of the iteration steps, but it does not include a clearly labeled 'Pseudocode' or 'Algorithm' block with structured, code-like formatting. |
| Open Source Code | No | The paper does not contain any explicit statement about the release of source code for the methodology described, nor does it provide a link to a code repository. |
| Open Datasets | No | Data models We use two different scenarios. The first is Gaussian mixture model in two dimensions. ... The second scenario is Gaussian model in higher dimensions. The Gaussian data was white, and the dimension took the values 5,10,100. This describes synthetic data generated for the experiments and does not refer to a publicly available dataset with concrete access information. |
| Dataset Splits | No | In the main results (bias removal), the last 30% of the data was analyzed as samples of the distribution; it is assumed here that this 30% represents the final distribution and all the algorithms have converged. This refers to the analysis of the generated MCMC samples, not a split of an input dataset for training, validation, or testing. |
| Hardware Specification | No | The paper does not contain any specific details about the hardware used to run the experiments, such as GPU models, CPU types, or memory specifications. |
| Software Dependencies | No | The paper does not specify any software libraries, frameworks, or their version numbers used in the implementation or for conducting experiments. |
| Experiment Setup | Yes | Step sizes are chosen as follows. According to the theory above, the step size for the proposed method is implicitly given based on the noise level as in (14). For the baseline Langevin methods, we use that same step size if nothing is mentioned, but conduct additional simulations with another step size which is 1/4 of the above. ... Each algorithm was run for 1,000,000 steps. ... The number of kernels (components) is varied from 1 to 4; ... We consider two different noise levels in the score function, a higher one (σ2 = 0.3) and a lower one (σ2 = 0.1). ... the dimension took the values 5,10,100. |