Undirected Graphical Models as Approximate Posteriors

Authors: Arash Vahdat, Evgeny Andriyash, William Macready

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results are provided in Sec. 4 on VAEs (Sec. 4.1), importance-weighted VAEs (Sec. 4.2), and structured prediction (Sec. 4.3), where we observe consistent improvement using UGMs.
Researcher Affiliation Industry 1NVIDIA, USA 2Sanctuary AI, Canada, Work done at D-Wave Systems. Correspondence to: Arash Vahdat <avahdat@nvidia.com>.
Pseudocode Yes Algorithm 1 summarizes training of DVAE##.
Open Source Code Yes Our implementation is available here.
Open Datasets Yes binarized MNIST (Salakhutdinov & Murray, 2008) and OMNIGLOT (Lake et al., 2015) datasets.
Dataset Splits No No explicit train/validation/test dataset splits were provided. While a 'test set' is mentioned, no details on validation or the partitioning of the full dataset are given.
Hardware Specification No No specific hardware details (such as CPU/GPU models, memory, or specific cloud instance types) were mentioned for running experiments.
Software Dependencies No No specific software dependencies with version numbers were provided. The paper mentions 'Tensorflow' and 'Qu PA library' but without version details.
Experiment Setup Yes p(x|z) is represented using a fully-connected neural network having two 200-unit hidden layers, tanh activations, and batch normalization similar to DVAE++ (Vahdat et al., 2018b), DVAE# (Vahdat et al., 2018a), and Gum Bolt (Khoshaman & Amin, 2018). Training DVAE## is done using Algorithm 1 with s = 10 and t = 1 using a piece-wise linear relaxation (Andriyash et al., 2018) for relaxed Gibbs samples. We follow (Vahdat et al., 2018a) for batch size, learning rate schedule, and KL warm up parameters.