Score-Based Generative Models Detect Manifolds
Authors: Jakiw Pidstrigach
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Score-based generative models, also called diffusion models ([Sohl-Dickstein et al., 2015, Song and Ermon, 2019, Song et al., 2021b, Vahdat et al., 2021]) and the related models ([Bordes et al., 2017, Ho et al., 2020, Kingma et al., 2021]) have shown great empirical success in many areas, such as image generation ([Jolicoeur-Martineau et al., 2021, Nichol and Dhariwal, 2021, Dhariwal and Nichol, 2021, Ho et al., 2022]), audio generation ([Chen et al., 2021, Kong et al., 2021, Jeong et al., 2021, Popov et al., 2021]) as well as in other applications ([Batzolis et al., 2021, De Bortoli et al., 2021, Zhou et al., 2021, Cai et al., 2020, Luo and Hu, 2021, Meng et al., 2021, Saharia et al., 2021, Li et al., 2022, Sasaki et al., 2021]). |
| Researcher Affiliation | Academia | Jakiw Pidstrigach Institut für Mathematik Universität Potsdam Karl-Liebknecht-Str. 24/25 14476 Potsdam pidstrigach@uni-potsdam.de |
| Pseudocode | No | The paper contains mathematical equations and theoretical derivations but no explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] |
| Open Datasets | Yes | We evaluated N from (10) on CIFAR-10, once for the difference between the sθ(x, t) from Song et al. [2021b] and log ˆpt, and once by just using a perturbed drift with a constant error, s(x, t) = log ˆpt + 1 2(1, 1, . . . , 1), see Figure 4. |
| Dataset Splits | No | The paper mentions using |
| Hardware Specification | Yes | This took less than 30 minutes on a single CPU. The experiments have been run on a single NVIDIA Quadro RTX 6000. |
| Software Dependencies | No | The paper mentions |
| Experiment Setup | Yes | The results in Figure 1 have been computed using 50000 2D samples from a two component Gaussian mixture. The numerical experiments for Figures 3 and 4 have been computed on CIFAR-10 with 50000 training images. We trained 50000 single Gaussian distributions on all 50000 images on CIFAR-10. |