Variational Inference for Monte Carlo Objectives
Authors: Andriy Mnih, Danilo Rezende
ICML 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the effectiveness of the proposed approach at training models for generative modelling and structured output prediction. We chose these two tasks because they involve models with hundreds of latent variables, which poses formidable challenges when estimating the gradients for the proposal distributions. In both cases we compare the performance of the VIMCO estimator to that of the NVIL estimator as well as to an effective biased estimator from the literature. We experiment with varying the number of samples in the objective to see how that affects the performance of the resulting models when using different estimators. The details of the training procedure are given in the supplementary material. Our first comparison is on the MNIST dataset of 28 28 images of handwritten digits, using the binarization of Salakhutdinov & Murray (2008) and the standard 50000/10000/10000 split into the training, validation, and test sets. Figure 1 shows the evolution of the training objective on the validation set as training proceeds. Table 1. Estimates of the negative log-likelihood (in nats) for generative modelling on MNIST. |
| Researcher Affiliation | Industry | Andriy Mnih AMNIH@GOOGLE.COM Danilo J. Rezende DANILOR@GOOGLE.COM Google Deep Mind |
| Pseudocode | Yes | The pseudocode for computing it is provided in the supplementary material. |
| Open Source Code | No | The paper mentions that 'The pseudocode for computing it is provided in the supplementary material.' but does not include any explicit statements about releasing source code or provide links to a code repository for the implemented methodology. (Explanation: No explicit statement about source code availability or a repository link is provided.) |
| Open Datasets | Yes | Our first comparison is on the MNIST dataset of 28 28 images of handwritten digits, using the binarization of Salakhutdinov & Murray (2008) and the standard 50000/10000/10000 split into the training, validation, and test sets. |
| Dataset Splits | Yes | Our first comparison is on the MNIST dataset of 28 28 images of handwritten digits, using the binarization of Salakhutdinov & Murray (2008) and the standard 50000/10000/10000 split into the training, validation, and test sets. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware (e.g., GPU models, CPU types, memory) used for running the experiments. (Explanation: No specific hardware details are mentioned.) |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers (e.g., 'PyTorch 1.9', 'Python 3.8'). (Explanation: No specific software dependencies with version numbers are mentioned.) |
| Experiment Setup | No | The paper mentions model architectures (e.g., 'SBN with three hidden layers of 200 binary latent variables (200-200-200-768)') and that 'The details of the training procedure are given in the supplementary material.' However, it does not provide concrete hyperparameter values (e.g., learning rate, batch size) or other system-level training settings directly in the main text. (Explanation: Specific hyperparameters and detailed training configurations are stated to be in the supplementary material, not the main text.) |