PAC-Bayesian Generalization Bounds for Adversarial Generative Models
Authors: Sokhna Diarra Mbacke, Florence Clerc, Pascal Germain
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Although our work is mainly theoretical, we perform numerical experiments showing non-vacuous generalization bounds for Wasserstein GANs on synthetic datasets. |
| Researcher Affiliation | Academia | 1Universit e Laval 2Mc Gill University. |
| Pseudocode | No | The paper describes proof ideas and theoretical derivations but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain an explicit statement about releasing its source code or a direct link to the implementation of its methodology. |
| Open Datasets | Yes | We perform experiments on two synthetic datasets: a mixture of 8 Gaussians arranged on a ring, and a mixture of 25 Gaussians arranged on a grid. These are standard synthetic datasets for GAN experiments, see, e.g, Dumoulin et al. (2017); Srivastava et al. (2017); Dieng et al. (2019)... We performed preliminary experiments on the MNIST dataset (Deng, 2012) |
| Dataset Splits | No | The paper mentions 'training and the test sets' for evaluation, but does not explicitly describe a separate validation set or provide details on dataset splits (e.g., percentages or counts for train/val/test). |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU/GPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions using the 'Pytorch-ignite library (Fomin et al., 2020)' and 'Pytorch' but does not specify version numbers for these or other software dependencies. |
| Experiment Setup | Yes | We performed a sweep over the values σ0 {10 7, 10 6, 10 5, 0.0001, 0.001, 0.01, 0.1}, and fix the hyperparameter λ = n 1024, where n is the size of the training set. ... In our chosen models, both the generator and critic are fully connected networks, and we use the Bj ork orthonormalization algorithm (Bj orck & Bowie, 1971) to enforce Lipschitz continuity on the critic. We performed experiments using both Re LU and Group Sort activations (Anil et al., 2019), and we report the results using Group Sort as it leads to more stability. |