InfoVAE: Balancing Learning and Inference in Variational Autoencoders
Authors: Shengjia Zhao, Jiaming Song, Stefano Ermon5885-5892
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive qualitative and quantitative analyses, we demonstrate that our models outperform competing approaches on multiple performance metrics. Experiments Variance Overestimation with ELBO training We first perform some simple experiments on toy data and MNIST to demonstrate that ELBO suffers from inaccurate inference in practice, and adding the scaling term λ in Eq.(5) Comprehensive Comparison In this section, we perform extensive qualitative and quantitative experiments on the binarized MNIST dataset to eval-uate the performance of different models. Table 1: Log likelihood estimates for different models on the MNIST dataset. Figure 6: Comparison of numerical performance. |
| Researcher Affiliation | Academia | Shengjia Zhao, Jiaming Song, Stefano Ermon Stanford University sjzhao,tsong,ermon@stanford.edu |
| Pseudocode | No | The paper does not contain any pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any explicit statement about releasing source code or a link to a code repository. |
| Open Datasets | Yes | We train ELBO and Info VAE (with MMD regularization) on binarized MNIST with different training set sizes ranging from 500 to 50000 images; We use the DCGAN architecture (Radford, Metz, and Chintala 2015) for both models. To be consistent with previous results, we use the stochastically binarized MNIST (Salakhutdinov and Murray 2008). |
| Dataset Splits | No | The paper mentions 'training set sizes' and the use of a 'test set', but it does not specify explicit percentages or counts for training, validation, and test splits, nor does it provide a detailed methodology for splitting the data to ensure reproducibility of the partitioning. |
| Hardware Specification | No | The paper does not explicitly describe any specific hardware (e.g., GPU or CPU models, memory specifications) used for running its experiments. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x, TensorFlow 2.x). |
| Experiment Setup | Yes | For Info VAE we choose the scaling coefficient λ = 500, information preference α = 0, and divergence optimized by MMD. For Info VAE, we use the scaling coefficient λ = 1000, and information preference α = 0. We choose the number of latent features dimension(Z) = 2 to plot the latent space, and 10 for all other experiments. We use a three layer deep network with 200 hidden units in each layer to simulate the highly flexible function family. In this setting we also use a highly flexible Pixel CNN as the decoder pθ(x|z) so that information preference is also a concern. Detailed experimental setup is explained in the Appendix. |