The Implicit Metropolis-Hastings Algorithm
Authors: Kirill Neklyudov, Evgenii Egorov, Dmitry P. Vetrov
NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | On real-world datasets (CIFAR-10, Celeb A, Image Net), we validate our approach using different deep generative models as independent proposals: DCGAN (Radford et al., 2015); Wasserstein GAN with gradient penalty (Gulrajani et al., 2017); variational auto-encoder (Kingma & Welling, 2014); Big GAN (Brock et al., 2018); MMD-GAN (Li et al., 2017). Every model is learned independently optimizing its original objective, what allows us to test the algorithm on a wide range of different proposals. For every proposal, we learn the discriminator from scratch (except Big GAN) and observe monotonous improvements of metrics throughout the learning. Further, we construct a Markov proposal by traversing the latent space of WPGAN generator via a Markov chain. Our experiments demonstrate that this proposal compares favorably against the independent proposal while using the same generator network. We present the empirical evaluation of the proposed algorithm and theory for both independent and Markov proposals. For independent proposals, we validate our theoretical result by demonstrating monotonous improvements of the sampling procedure throughout the learning of the discriminator. |
| Researcher Affiliation | Collaboration | Kirill Neklyudov Samsung-HSE Laboratory HSE , Moscow, Russia Samsung AI Center Moscow k.necludov@gmail.com Evgenii Egorov Skoltech , Moscow, Russia egorov.evgenyy@ya.ru Dmitry Vetrov Samsung-HSE Laboratory HSE , Moscow, Russia Samsung AI Center Moscow vetrovd@yandex.ru National Research University Higher School of Economics Skolkovo Institute of Science and Technology |
| Pseudocode | Yes | Algorithm 1 The Metropolis-Hastings algorithm; Algorithm 2 Metropolis-Hastings GAN; Algorithm 3 The implicit Metropolis-Hastings algorithm |
| Open Source Code | Yes | Code reproducing all experiments is available online3. 3https://github.com/necludov/implicit-MH |
| Open Datasets | Yes | On real-world datasets (CIFAR-10, Celeb A, Image Net), we validate our approach using different deep generative models as independent proposals: DCGAN (Radford et al., 2015); Wasserstein GAN with gradient penalty (Gulrajani et al., 2017); variational auto-encoder (Kingma & Welling, 2014); Big GAN (Brock et al., 2018); MMD-GAN (Li et al., 2017). We present the empirical evaluation of the proposed algorithm and theory for both independent and Markov proposals. In the experimental part of the paper, we empirically validate the proposed algorithm on the real-world datasets (CIFAR-10, Celeb A, Image Net) using different generative models as proposals. |
| Dataset Splits | No | The paper mentions using 10k samples for metric evaluation, but it does not specify any explicit training, validation, or test dataset splits (e.g., percentages, counts, or references to predefined splits). |
| Hardware Specification | No | The paper does not provide any specific details about the hardware used for running the experiments (e.g., GPU models, CPU types, memory). |
| Software Dependencies | No | The paper mentions 'Py Torch version of the Inception V3 network (Paszke et al., 2017)' but does not provide specific version numbers for PyTorch or other key software components used in the experiments. |
| Experiment Setup | Yes | For GANs, we take the generator, for VAE, we take the decoder and the prior. Then we learn the discriminator from scratch for all models (except Big GAN; there we finetune the head of the discriminator) and monitor the performance of the Algorithm 3 with iterations. In our experiments, we take t = π/3. For loss estimation, we condition samples from the proposal on samples from the dataset x q(x | y), y p(y). |