A Contrastive Learning Approach for Training Variational Autoencoder Priors

Authors: Jyoti Aneja, Alex Schwing, Jan Kautz, Arash Vahdat

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments confirm that the proposed noise contrastive priors improve the generative performance of state-of-the-art VAEs by a large margin on the MNIST, CIFAR-10, Celeb A 64, and Celeb A HQ 256 datasets.
Researcher Affiliation Collaboration 1University of Illinois at Urbana-Champaign, 2NVIDIA 1{janeja2, aschwing}@illinois.edu, 2{jkautz,avahdat}@nvidia.com
Pseudocode No The paper does not contain any pseudocode or clearly labeled algorithm blocks.
Open Source Code No The paper states, 'We borrow the exact training setup from [17] and implement our method using their publicly available code.3' with a footnote to 'https://github.com/ParthaEth/Regularized_autoencoders-RAE-'. This suggests using a third-party's code for implementation, not that the authors' specific implementation of NCP is publicly released or available at this link.
Open Datasets Yes Our experiments confirm that the proposed noise contrastive priors improve the generative performance of state-of-the-art VAEs by a large margin on the MNIST, CIFAR-10, Celeb A 64, and Celeb A HQ 256 datasets.
Dataset Splits No The paper mentions using training and test sets but does not provide specific percentages or sample counts for training, validation, and test splits required for reproduction.
Hardware Specification No The paper mentions receiving 'compute support' from 'the NGC team at NVIDIA' in the acknowledgements, but does not provide specific details on the GPU or CPU models, memory, or any other hardware specifications used for the experiments.
Software Dependencies No The paper does not provide specific version numbers for any software dependencies or libraries used in the experiments.
Experiment Setup Yes For generating samples from the model, we use SIR with 5K proposal samples. To report log-likelihood results, we train models with small latent spaces on the dynamically binarized MNIST [46] dataset. ... We adopt the common practice of reducing the temperature of the base prior p(z) by scaling down the standard-deviation of the conditional Normal distributions [38]. ... Similarly, we achieve diverse, high-quality images by re-adjusting the BN statistics as described by [74]. ... In Tab. 8, we observe that increasing both the number of proposal samples in SIR and the LD iterations leads to a noticeable improvement in FID score.