GAN-EM: GAN Based EM Learning Framework
Authors: Wentian Zhao, Shaojie Wang, Zhihuai Xie, Jing Shi, Chenliang Xu
IJCAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To evaluate our model, we perform the clustering task based on MNIST and achieve lowest error rate with all 3 different numbers of clusters: 10, 20 and 30, which are common settings in previous works. We also test the semi-supervised classification performance on MNIST and SVHN with partially labeled data, both results being rather competitive compared to recently proposed generative models. Especially, on SVHN dataset, GAN-EM outperforms all other models. Apart from the two commonly used datasets, we test our model on an additional dataset, Celeb A, under both unsupervised and semi-supervised settings, which is a more challenging task because attributes of human faces are rather abstract. It turns out that our model still achieves the best results. |
| Researcher Affiliation | Academia | 1University of Rochester 2Tsinghua University |
| Pseudocode | Yes | Algorithm 1 GAN-EM |
| Open Source Code | Yes | Supplementary material can be found at http://www.cs. rochester.edu/ cxu22/p/. |
| Open Datasets | Yes | We perform unsupervised clustering on MNIST [Le Cun et al., 1989] and Celeb A [Liu et al., 2015] datasets, and semisupervised classification on MNIST, SVHN [Netzer et al., 2011] and Celeb A datasets. |
| Dataset Splits | No | The paper mentions evaluating on datasets and discusses 'test error rate' but does not explicitly provide specific training/validation/test dataset splits (e.g., percentages or exact sample counts for each split) in the main text. |
| Hardware Specification | No | The paper does not specify the hardware used (e.g., GPU models, CPU types) for running experiments. |
| Software Dependencies | No | The paper mentions 'RMSprop optimizer' but does not specify software dependencies with version numbers (e.g., specific Python, PyTorch, or CUDA versions). |
| Experiment Setup | Yes | We apply RMSprop optimizer to all 3 networks G, D and E with learning rate 0.0002 (decay rate: 0.98). ... In each M-step, there are 5 epoches with a minibatch size of 64 for both the generated batch and the real samples batch. We use a same update frequency for generator and discriminator. For E-step, we generate samples using well trained generator with batch size of 256, then we apply 1000 iterations to update E-net. |