Multiscale Invertible Generative Networks for High-Dimensional Bayesian Inference

Authors: Shumao Zhang, Pengchuan Zhang, Thomas Y Hou

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We study two high-d Bayesian inverse problems (BIPs) in Section 6.1 as test beds for distribution approximation and multi-mode capture. We also apply Ms IGN to the image synthesis task to benchmark with flow-based generative models and demonstrate its interpret-ability in Section 6.2.
Researcher Affiliation Collaboration 1Department of Computational & Mathematical Sciences, Caltech, Pasadena, California, USA 2MSR AI Lab, Redmond, Washington, USA.
Pseudocode Yes Algorithm 1 Train Ms IGN by optimizing the Jeffreys divergence in a multi-stage manner
Open Source Code No The paper does not contain any explicit statement about making the source code available or provide a link to a code repository.
Open Datasets Yes Table 2. Bits-per-dimension value comparison with baseline models of flow-based generative networks. All models in this table do not use the variational dequantization technique in (Ho et al., 2019). *: Score obtained by our own reproducing experiment. MODEL MNIST CIFAR-10 CELEBA 64 IMAGENET 32 IMAGENET 64
Dataset Splits No The paper mentions using datasets like MNIST, CIFAR-10, CelebA, and ImageNet for image synthesis, but it does not specify the exact training, validation, and test split percentages or sample counts in the provided text. It refers to Appendix F for more details, which is not available.
Hardware Specification No The paper does not provide specific details about the hardware used for running experiments, such as GPU or CPU models.
Software Dependencies No The paper refers to using 'the invertible block of Glow (Kingma & Dhariwal, 2018)' and mentions 'Adam' for optimization, but it does not specify version numbers for any software dependencies or libraries.
Experiment Setup No The paper states high-level architectural details like 'L = 6 scales' and 'average pooling with kernel size 2 and stride 2' and mentions referring to Appendix F for 'More details of experimental setting', but does not explicitly provide specific hyperparameter values (e.g., learning rate, batch size, epochs) in the main text.