Structured Uncertainty in the Observation Space of Variational Autoencoders
Authors: James Langley, Miguel Monteiro, Charles Jones, Nick Pawlowski, Ben Glocker
TMLR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform the comparison in two datasets: the CELEBA dataset (Liu et al., 2015) and the UK Biobank (UKBB) Brain Imaging dataset (Miller et al., 2016). For all models, we use a latent space of dimension 128. For the low-rank model, we use a rank of 25. [...] Quantitative evaluation of generative modelling is an inherently difficult task due to its subjective nature. [...] The Fréchet Inception Distance (FID) (Heusel et al., 2017) is the current standard choice of metric due to its consistency with human perception (Borji, 2019). We use it to evaluate our generative models (Seitzer, 2020) and report the results in Table 1. |
| Researcher Affiliation | Collaboration | James Langley EMAIL Department of Computing, Imperial College London, UK Miguel Monteiro EMAIL Department of Computing, Imperial College London, UK Charles Jones EMAIL Department of Computing, Imperial College London, UK Nick Pawlowski EMAIL Microsoft Research, Cambridge, UK Department of Computing, Imperial College London, UK Ben Glocker EMAIL Department of Computing, Imperial College London, UK |
| Pseudocode | No | The paper describes methods and equations for the SOS-VAE, but does not present any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | All code is made publicly available on https://github.com/biomedia-mira/sos-vae together with clear instructions how to fully reproduce our results on CELEBA to facilitate future work and ease of comparisons. |
| Open Datasets | Yes | We perform the comparison in two datasets: the CELEBA dataset (Liu et al., 2015) and the UK Biobank (UKBB) Brain Imaging dataset (Miller et al., 2016). |
| Dataset Splits | No | The paper mentions using the CELEBA dataset (Liu et al., 2015) and the UK Biobank (UKBB) Brain Imaging dataset (Miller et al., 2016) but does not provide specific details on how these datasets were split into training, validation, and test sets. |
| Hardware Specification | Yes | In contrast, the SOS-VAE model used in our experiments is trained on a single 16-GB T4 GPU in 72 hours, which is a substantial difference of practical importance. |
| Software Dependencies | No | The paper does not explicitly list specific software dependencies with version numbers for its implementation. While PyTorch is mentioned in the context of a tool for FID calculation, it does not specify the PyTorch version used for the main model implementation. |
| Experiment Setup | Yes | For all models, we use a latent space of dimension 128. For the low-rank model, we use a rank of 25. For the CELEBA dataset we use a target KL loss , ξKL, of 45 for both models and ξH = 504750 for our model. For the UKBB dataset we use a target KL loss, ξKL, of 15 for both models and ξH = 198906 for our model. |