Denoising Diffusion Variational Inference: Diffusion Models as Expressive Variational Posteriors

Authors: Wasu Top Piriyakulkij, Yingheng Wang, Volodymyr Kuleshov

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate DDVI on synthetic benchmarks and on a real problem in biological data analysis inferring human ancestry from genetic data. Our method outperforms strong baselines on the Thousand Genomes dataset (Siva 2008) and learns a low-dimensional latent space that preserves biologically meaningful structure (Haghverdi, Buettner, and Theis 2015). We compare DDVI with Auto-Encoding Variational Bayes (AEVB) (Kingma and Welling 2013), AEVB with inverse autoregressive flow posteriors (AEVB-IAF) (Kingma et al. 2016), Adversarial Auto-Encoding Bayes (AAEB) (Makhzani et al. 2015), and Path Integral Sampler (PIS) (Zhang and Chen 2021) on MNIST (Lecun et al. 1998) and CIFAR-10 (Krizhevsky and Hinton 2009) in unsupervised and semi-supervised learning settings, and also on the Thousand Genomes dataset (Siva 2008). From Table 1 and Table 7 in Appendix, we see our method DDVI achieve best ELBO in all but one scenario, in which it still performs competitively.
Researcher Affiliation Academia Wasu Top Piriyakulkij*1, Yingheng Wang*1, Volodymyr Kuleshov1,2 1Department of Computer Science, Cornell University 2The Jacobs Technion-Cornell Institute, Cornell Tech EMAIL
Pseudocode No The paper describes the optimization steps in Section 3.4 'Optimization: Extending Wake-Sleep' by listing steps 1 and 2, but these are presented as prose within the text rather than as a formally structured pseudocode or algorithm block.
Open Source Code No The paper does not contain any explicit statements about releasing source code, nor does it provide links to a code repository.
Open Datasets Yes We compare DDVI with Auto-Encoding Variational Bayes (AEVB) (Kingma and Welling 2013), AEVB with inverse autoregressive flow posteriors (AEVB-IAF) (Kingma et al. 2016), Adversarial Auto-Encoding Bayes (AAEB) (Makhzani et al. 2015), and Path Integral Sampler (PIS) (Zhang and Chen 2021) on MNIST (Lecun et al. 1998) and CIFAR-10 (Krizhevsky and Hinton 2009) in unsupervised and semi-supervised learning settings, and also on the Thousand Genomes dataset (Siva 2008).
Dataset Splits No The paper mentions specific numbers of labels observed for semi-supervised learning (e.g., '1,000 for MNIST and 10,000 for CIFAR-10'), but it does not provide explicit train/test/validation split percentages or sample counts for the full datasets in the main text.
Hardware Specification No The paper does not provide any specific hardware details such as GPU or CPU models used for running experiments.
Software Dependencies No The paper does not specify any software dependencies with version numbers.
Experiment Setup No The priors, model architecture, and training details can also be founded in Appendix H.