Ai-sampler: Adversarial Learning of Markov kernels with involutive maps
Authors: Evgenii Egorov, Riccardo Valperga, Stratis Gavves
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, we use a bootstrap process to learn to sample from a given analytic density and show, on various synthetic, and real-world examples, that the Markov chain resulting from the learned transition kernel produces samples from the target density much more effectively and efficiently than other existing methods. |
| Researcher Affiliation | Academia | 1Amsterdam Machine Learning Lab, University of Amsterdam, the Netherlands 2Riccardo Valperga VISLab, University of Amsterdam, the Netherlands. |
| Pseudocode | Yes | The bootstrap process is outlined in Algorithm 1. |
| Open Source Code | Yes | We provide code for reproducing the experiments at https://github.com/ricvalp/aisampler. |
| Open Datasets | Yes | We use the same posterior distribution they used, resulting from a Bayesian probabilistic logistic regression model on three famous datasets: heart (14 covariates, 532 data points), australian (15 covariates, 690 data points), and german (25 covariates, 1000 data points). |
| Dataset Splits | No | The paper mentions splitting data into "train and test subsets" and using "burn-in steps" for the Markov chain, but it does not explicitly describe a distinct "validation" dataset split with percentages or counts for hyperparameter tuning or model selection. |
| Hardware Specification | Yes | The models have been trained on one NVIDIA A100. For more more than 104 parallel chains, the jitting time becomes prohibitively long and therefore not worthy. (Figure 5 mentions RTX3090 GPU) |
| Software Dependencies | No | The paper mentions implementing the Ai-sampler and HMC in JAX (Bradbury et al., 2018) and Flax (Heek et al., 2023), but it does not specify explicit version numbers for these software packages, which is required for reproducibility. |
| Experiment Setup | Yes | For experiments with 2D distributions we use H enon maps with the function V being a two-layer MLP with hidden dimension 32 and compose 5 of such layers to construct the function gθ from Eq. (6). To sample from the Bayesian logistic regression posterior we set the hidden dimension of the two-layer MLP is 64. For the discriminator we use the product parameterization using two three-layer MLP with hidden dimensions 32 for the experiments with the 2D distribution, and 128 for the Bayesian posterior. |