On the Posterior Distribution in Denoising: Application to Uncertainty Quantification
Authors: Hila Manor, Tomer Michaeli
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Code and examples are available on the project website. We illustrate our approach with several pre-trained denoisers on multiple domains, showing its practical benefit in uncertainty visualization. We conduct experiments with our proposed approach for uncertainty visualization and marginal posterior distribution estimation on additional real data in multiple domains using different models. |
| Researcher Affiliation | Academia | Hila Manor Faculty of Electrical and Computer Engineering Technion Israel Institute of Technology hila.manor@campus.technion.ac.il Tomer Michaeli Faculty of Electrical and Computer Engineering Technion Israel Institute of Technology tomer.m@ee.technion.ac.il |
| Pseudocode | Yes | Algorithm 1 Efficient computation of posterior principal components |
| Open Source Code | Yes | Code and examples are available on the project website. As part of the ongoing effort to make the field of deep learning more reproducible and open, we publish our code at https://hilamanor.github.io/Gaussian Denoising Posterior/. |
| Open Datasets | Yes | For natural images, we use Swin IR (Liang et al., 2021) that was pre-trained on 800 DIV2K (Agustsson & Timofte, 2017) images, 2650 Flickr2k (Lim et al., 2017) images, 400 BSD500 (Arbelaez et al., 2010) images and 4,744 WED (Ma et al., 2016) images. We showcase examples from the Celeb AMask-HQ dataset (Lee et al., 2020). Figure 4 illustrates the approach on a handwritten digit from the MNIST (Le Cun, 1998) dataset. For the medical and microscopy domain we use Noise2Void (Krull et al., 2019), trained and tested for blind-denoising on the FMD dataset (Zhang et al., 2019). |
| Dataset Splits | No | The paper mentions datasets used for pre-trained models and describes training a simple neural network and a CNN, but it does not specify explicit training/validation/test dataset splits or cross-validation setup for reproducibility of its own training processes. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments. It mentions memory footprint but not the hardware producing it. |
| Software Dependencies | No | The paper mentions some software components like Adam (Kingma & Ba, 2015) and Si LU (Hendrycks & Gimpel, 2016), but it does not provide specific version numbers for these or other key software libraries and frameworks necessary for replication. |
| Experiment Setup | Yes | Algorithm 1 requires three hyper-parameters as input. The first is the small constant c, which is used for the linear approximation in (15). The second is N, which is the number of principal components we seek. The last is K, which is the number of iterations to preform. In all our experiments we used c = 10 5 and N = 3. For the N2V experiments we used K = 100 while for the rest we used K = 50. This denoiser comprises 5 layers with (100, 200, 200, 100) hidden features and Si LU (Hendrycks & Gimpel, 2016) activation units. We trained the network using Adam (Kingma & Ba, 2015) for 300 epochs, with a learning rate of 0.005. Here, we train and use a simple CNN with 10 layers of 64 channels, separated by Re LU activation layers followed by batch normalization layers. |