Differentiable Sampling of Categorical Distributions Using the CatLog-Derivative Trick
Authors: Lennert De Smet, Emanuele Sansone, Pedro Zuidberg Dos Martires
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section we evaluate the performance of Inde Cate R on a set of standard benchmarks from the literature. |
| Researcher Affiliation | Academia | Lennert De Smet KU Leuven Emanuele Sansone KU Leuven Pedro Zuidberg Dos Martires Örebro University |
| Pseudocode | No | The paper describes methods and proofs using mathematical notation and prose, but it does not contain any structured pseudocode or algorithm blocks with explicit labels like 'Algorithm' or 'Pseudocode'. |
| Open Source Code | Yes | Our code is publicly available at https://github.com/ML-KULeuven/catlog. |
| Open Datasets | Yes | We optimise the DVAE on the three main datasets from the literature, being MNIST [21], F-MNIST [42] and Omniglot [20]. |
| Dataset Splits | Yes | We optimise the DVAE on the three main datasets from the literature, being MNIST [21], F-MNIST [42] and Omniglot [20]. and As evaluation metrics, we show the negated training and test set ELBO in combination with the variance of the gradients throughout training. These datasets have standard, predefined train/test splits commonly used in the community. |
| Hardware Specification | No | The paper states, 'The different methods were benchmarked using discrete GPUs and every method had access to the same computational resources.' However, it does not provide specific hardware details such as GPU models, CPU models, or memory specifications in the main text. It mentions these details are in the appendix, but they are not provided in the main paper. |
| Software Dependencies | No | The paper mentions that 'We implemented Inde Cate R and RLOO in TensorFlow [1]', but it does not specify a version number for TensorFlow or any other software dependencies needed for replication. |
| Experiment Setup | Yes | We performed a hyperparameter search for the learning rate and the temperature of GS-F. Parameters were optimised using RMSProp. (from Figure 2 caption) and Inde Cate R and RLOO-S both use 2 samples, while RLOO-F and Gumbel-Softmax (GS) use 800 samples. (from Figure 2 caption). |