Cooperative Learning of Energy-Based Model and Latent Variable Model via MCMC Teaching
Authors: Jianwen Xie, Yang Lu, Ruiqi Gao, Ying Nian Wu
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments show that the cooperative learning algorithm can learn realistic models of images. |
| Researcher Affiliation | Collaboration | Jianwen Xie,1,2 Yang Lu,1,3 Ruiqi Gao,1 Ying Nian Wu1 1Department of Statistics, University of California, Los Angeles, USA 2Hikvision Research America 3Amazon RSML (Retail System Machine Learning) Group |
| Pseudocode | Yes | Algorithm 1 Coop Nets Algorithm |
| Open Source Code | Yes | The code and more results can be found at http://www.stat. ucla.edu/~ywu/Coop Nets/main.html |
| Open Datasets | Yes | We conduct experiments on synthesizing images of categories from Imagenet ILSVRC2012 dataset (Deng et al. 2009). ... The training data are 10,000 human faces randomly sampled from Celeb A dataset (Liu et al. 2015). ... We also apply the same method to MIT forest road category and hotel room category (Zhou et al. 2014). |
| Dataset Splits | No | The paper specifies training image counts for different categories (e.g., "The numbers of images sampled from each category are 50, 100, 300, 500, 700, 900, and 1100 respectively") and uses "training images" and "testing images" in context. It also mentions "1,000 testing images" for pattern completion. However, it does not provide explicit training/validation/test dataset split percentages or specific counts for a validation set, nor does it refer to predefined splits that include a validation portion. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, RAM, or cloud computing instance types) used for running its experiments. |
| Software Dependencies | No | We use the Mat Conv Net of (Vedaldi and Lenc 2015) for coding. The paper names a software package but does not provide its version number or other specific software dependencies with versions. |
| Experiment Setup | Yes | We use lp = 20 or 30 steps of Langevin revision dynamics within each learning iteration, and the Langevin step size is set at 0.003. The learning rate is 0.01. ... The standard deviation of the noise vector is σ = 0.3. The learning rate is 10^-6. The number of generator learning steps is 1 at each cooperative learning iteration. We run 10^4 cooperative learning iterations to train the models. ... We set the number of Langevin dynamics steps in each learning iteration to lp=10 and the step size to 0.002. The learning rate is 0.07. The number of learning iterations is about 1,000. |