MIWAE: Deep Generative Modelling and Imputation of Incomplete Data Sets
Authors: Pierre-Alexandre Mattei, Jes Frellsen
ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We illustrate our approach by training a convolutional DLVM on incomplete static binarisations of MNIST. Moreover, on various continuous data sets, we show that MIWAE provides extremely accurate single imputations, and is highly competitive with state-of-the-art methods. |
| Researcher Affiliation | Academia | 1Department of Computer Science, IT University of Copenhagen, Denmark. Correspondence to: Pierre-Alexandre Mattei <pima@itu.dk>, Jes Frellsen <jefr@itu.dk>. |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. Methods are described in prose. |
| Open Source Code | No | The paper does not provide concrete access to source code for the methodology. There is no link or explicit statement of code release. |
| Open Datasets | Yes | We illustrate the features of MIWAE by training a DLVM on an incomplete version of the static binarisation of MNIST. We consider a simple setting: with 50% of the pixels missing uniformly at random (in a MCAR fashion). (Dua & Efi, 2017). URL http://archive.ics.uci.edu/ml. |
| Dataset Splits | Yes | To compare models, we evaluate estimates of their test log-likelihood obtained using importance sampling with 5000 samples and an inference network refitted on the test set, as suggested by Cremer et al. (2018) and Mattei & Frellsen (2018b). |
| Hardware Specification | No | The paper does not provide specific hardware details (GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments. |
| Software Dependencies | No | The paper does not provide specific ancillary software details with version numbers (e.g., library or solver names with version numbers like Python 3.8, CPLEX 12.4). |
| Experiment Setup | Yes | The intrinsic dimension d is fixed to 10, which may be larger than the actual number of features in the data, but DLVMs are known to automatically ignore some latent dimensions (Dai et al., 2018); both encoder and decoder are multi-layer perceptrons with 3 hidden layers (with 128 hidden units) and tanh activations; we use products of Student s t for the variational family (following Domke & Sheldon, 2018) and the observation model (following Takahashi et al., 2018). We perform 500 000 gradient steps for all data sets; no regularisation scheme is used, but the observation model is constrained so that the eigenvalues of its covariances are larger than 0.01 (as suggested by Mattei & Frellsen, 2018a). |