Recursive Inference for Variational Autoencoders
Authors: Minyoung Kim, Vladimir Pavlovic
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We test the proposed recursive inference model3 on several benchmark datasets. We highlight improved test likelihood scores and reduced inference time, compared to the semi-amortized VAEs. The test log-likelihood scores are summarized in Table 1. Overall the results indicate that our recursive mixture encoder (RME) outperforms the competing approaches consistently for all datasets. |
| Researcher Affiliation | Collaboration | Minyoung Kim1 1Samsung AI Center Cambridge, UK mikim21@gmail.com Vladimir Pavlovic1,2 2Rutgers University Piscataway, NJ, USA vladimir@cs.rutgers.edu |
| Pseudocode | Yes | Algorithm 1 Recursive Learning Algorithm for Mixture Inference Model. |
| Open Source Code | Yes | The code is publicly available from https://github.com/minyoungkim21/recmixvae |
| Open Datasets | Yes | Datasets. MNIST [18], OMNIGLOT [17], SVHN [25], and Celeb A [19]. |
| Dataset Splits | Yes | We follow train/test partitions provided in the data, where 10% of the training sets are randomly held out for validation. For Celeb A, we randomly split data into 80%/10%/10% train/validation/test sets. |
| Hardware Specification | Yes | All models are run on the same machine with a single GPU (RTX 2080 Ti), Core i7 3.50GHz CPU, and 128 GB RAM. |
| Software Dependencies | No | The paper does not provide specific ancillary software details with version numbers (e.g., library or solver names with version numbers). |
| Experiment Setup | Yes | We vary the latent dim(z), small (20) or large (50). To report the test log-likelihood scores log p(x), we use the importance weighted sampling estimation (IWAE) method [1] with 100 samples (Supplement for details). For each model/dataset, we perform 10 runs with different random train/validation splits, where each run consists of three trainings by starting with different random model parameters, among which only one model with the best validation result is chosen. Learning rate γ. KL bound C. Output: Learned inference and decoder models. C = 500.0 works well empirically. The mixture order M (the number of the inner iterations) is typically small as well (e.g., between 1 and 4). |