Adversarially-learned Inference via an Ensemble of Discrete Undirected Graphical Models
Authors: Adarsh Keshav Jeewajee, Leslie Kaelbling
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4 Experiments Our first experiment (4.1), shows that although undirected graphical models with empirical risk minimization (EGMs) are trained specifically for certain inference tasks, our adversarially-trained graphical models (AGMs) can perform comparatively, despite having never seen those tasks prior to training. The second experiment (4.2) is our main experiment which showcases the generalization capabilities of AGMs across unseen inference tasks on images. We also compare AGMs against state-of-the-art neural models Gibbs Net and VAEAC which, like AGMs, were designed for arbitrary conditioning. In the last experiment (4.3), we show that the combination of AGMs and their neural learner provide a viable alternative for sampling from joint probability distributions in one shot, compared to Gibbs samplers defined on EGMs. |
| Researcher Affiliation | Academia | Adarsh K. Jeewajee MIT CSAIL jaks19@mit.edu Leslie P. Kaelbling MIT CSAIL lpk@csail.mit.edu |
| Pseudocode | No | No pseudocode or algorithm blocks were found in the paper. |
| Open Source Code | No | The paper does not provide an explicit statement about releasing source code or a link to a code repository. |
| Open Datasets | Yes | Concerning data sets, we use: ten binary data sets used in previous probabilistic modelling work (example Gens and Pedro [2013]) spanning 16 to 1359 variables, two binarized image data sets (28x28 pixels) which are MNIST [Le Cun and Cortes, 2010] with digits 0 to 9, and Caltech-101 Silhouettes [Li et al., 2004] with shape outlines from 101 different categories, one discrete image data set (30x40 pixels) made of downsampled segmentation masks from the Stanford background dataset, and one RGB image data set (32x32 pixels) which is the SVHN dataset with digits of house numbers (see appendix for how we encode continuous RGB values through node marginals). |
| Dataset Splits | No | The paper mentions calibrating models on training data and testing on unseen points, but does not provide specific training/test/validation split percentages, absolute sample counts, or detailed splitting methodology. |
| Hardware Specification | No | The paper mentions that belief propagation runs on a GPU but does not provide specific hardware details such as GPU/CPU models, processor types, or memory amounts used for experiments. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers (e.g., library or solver names with versions). |
| Experiment Setup | Yes | We calibrate our models on each training data set separately, and test on 1000 unseen points. The inference task fractional(0.7) is used to test every model. EGMs train by minimizing the conditional log likelihood score under the inference task given by fractional(0.5). We use identical randomized edge sets of size 5|V| for non-image data, while a grid graph structure is used with image data (last four rows of table 1). We randomly sample M random vectors z1, . . . , z M from the standard multivariate Gaussian distribution... for j = 1, . . . , M. For the Gibbs sampler deļ¬ned on B, we try two scenarios: one where it uses no burn-in cycles... and one scenario where it has 10 burn-in cycles. |