Improving Variational Autoencoder Estimation from Incomplete Data with Mixture Variational Families

Authors: Vaidotas Simkus, Michael U. Gutmann

TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the proposed methods for VAE estimation on synthetic and realistic data sets with missing data (section 6).
Researcher Affiliation Academia Vaidotas Simkus EMAIL Michael U. Gutmann EMAIL School of Informatics University of Edinburgh
Pseudocode Yes Algorithm 1 Shared computation of the De Miss VAE learning objectives
Open Source Code Yes The methods are summarised in table 1 and the code implementation is available at https://github.com/ vsimkus/demiss-vae.
Open Datasets Yes We here evaluate the proposed methods on real-world data sets from the UCI repository (Dua & Graff, 2017; Papamakarios et al., 2017).
Dataset Splits No The paper mentions evaluating on a 'complete test data set' and a '20K sample data set used to fit the VAEs' but does not provide specific split percentages or counts (e.g., train/validation/test splits). The missingness percentages (e.g., 20/50/80%) refer to data incompleteness, not dataset splits for training and evaluation.
Hardware Specification No The paper does not provide specific hardware details such as GPU models (e.g., NVIDIA A100), CPU models, or memory amounts used for running the experiments.
Software Dependencies No The paper mentions the 'AMSGrad optimiser (Reddi et al., 2018)' and 'STL gradients (Roeder et al., 2017)' which are algorithms/techniques, but it does not specify software libraries with version numbers (e.g., PyTorch 1.9, TensorFlow 2.x).
Experiment Setup Yes We then fitted a VAE model with 2-dimensional latent space using diagonal Gaussian encoder and decoder distributions, and a fixed standard Normal prior. For the decoder and encoder networks we used fullyconnected residual neural networks with 3 residual blocks, 200 hidden dimensions, and Re LU activations. To optimise the model parameters we have used AMSGrad optimiser (Reddi et al., 2018) with a learning rate of 10 3 for a total of 500 epochs.