Semantic uncertainty intervals for disentangled latent spaces

Authors: Swami Sankaranarayanan, Anastasios Angelopoulos, Stephen Bates, Yaniv Romano, Phillip Isola

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 3 Experiments3.1 Dataset descriptions3.2 Experimental setup3.3 Findings
Researcher Affiliation Academia 1MIT 2University of California, Berkeley 3Technion Israel Institute of Technology
Pseudocode Yes Algorithm 1 Quantile GAN encoder training
Open Source Code No The code will be released in the near future.
Open Datasets Yes FFHQ We use the Style GAN framework pretrained using the Flickr-Faces-HQ (FFHQ) dataset [25]. FFHQ is a publicly available dataset consisting of 70,000 high-quality images at 1024 1024 resolution. ... Celeb A-HQ We use the Celeb A-HQ dataset [23]... CLEVR dataset [22].
Dataset Splits Yes We generate 100k samples per model and generate a random 80-10-10 split for training, calibration and validation.
Hardware Specification No The paper's self-assessment in section 3d explicitly states: "Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [No]"
Software Dependencies No The paper mentions software like Style GAN2, ResNet-50, Ranger optimizer, VGG network, but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes The hyperparameter weights (Eq 8) are set to c1 = c2 = 10.0. ... a flat learning rate of 0.001 for all our experiments. ... The risk level α and the user-specified error threshold δ are fixed to 0.1, unless specified otherwise. ... For the image super-resolution training, we augment the input dataset by using different levels of downsampled inputs, i.e., we take the raw input and apply a random downsampling factor from {1, 4, 8, 16, 32} and resize it to the original dimensions.