Improved Variational Inference with Inverse Autoregressive Flow

Authors: Durk P. Kingma, Tim Salimans, Rafal Jozefowicz, Xi Chen, Ilya Sutskever, Max Welling

NeurIPS 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In experiments, we show that IAF significantly improves upon diagonal Gaussian approximate posteriors.
Researcher Affiliation Collaboration Diederik P. Kingma dpkingma@openai.com Tim Salimans tim@openai.com Rafal Jozefowicz rafal@openai.com Xi Chen peter@openai.com Ilya Sutskever ilya@openai.com Max Welling M.Welling@uva.nl University of Amsterdam, University of California Irvine, and the Canadian Institute for Advanced Research (CIFAR).
Pseudocode Yes Algorithm 1: Pseudo-code of an approximate posterior with Inverse Autoregressive Flow (IAF)
Open Source Code Yes Code for reproducing key empirical results is available online3.
Open Datasets Yes Table 1 shows results on MNIST for these types of posteriors. We also evaluated IAF on the CIFAR-10 dataset of natural images.
Dataset Splits No The paper mentions 'dynamically sampled binarized MNIST version used in previous publications' and 'Hugo Larochelle s statically binarized MNIST', but does not provide specific split percentages or sample counts for training, validation, and test sets. It relies on previously established datasets without detailing the splits within the paper.
Hardware Specification Yes Sampling took about 0.05 seconds/image with the Res Net VAE model, versus 52.0 seconds/image with the Pixel CNN model, on a NVIDIA Titan X GPU.
Software Dependencies No The paper does not provide specific software details with version numbers (e.g., Python 3.x, PyTorch 1.x) that would be needed to replicate the experiment.
Experiment Setup Yes Please see appendix C for details on the architectures of the generative model and inference models.