A Solvable High-Dimensional Model of GAN
Authors: Chuang Wang, Hong Hu, Yue Lu
NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Numerical verification. We verify the theoretical prediction given by the ODE (8) via numerical simulations under the settings stated in Example 1. The results are shown in Figure 1. The number of features is d = 2, and ck and eck are both Gaussian with zero mean and covariance diag([5, 3]). The dimension is n = 5, 000, and the learning rates of the generator and discriminator are eτ = 0.04 and τ = 0.2 respectively. After testing different noise strength ηT = ηG = 2, 1, 4, we have observed at least three nontrivial dynamical patterns: success, oscillating or mode collapsing. In all these experiments, our theoretical predictions match the actual trajectories of the macroscopic states pretty well. |
| Researcher Affiliation | Academia | 1. National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Science, 95 Zhong Guan Cun Dong Lu, Beijing 100190, China 2. John A. Paulson School of Engineering and Applied Sciences, Harvard University 33 Oxford Street, Cambridge, MA 02138, USA |
| Pseudocode | No | The paper describes the training algorithm using mathematical equations like (5) for updates, but it does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any explicit statement or link about the availability of open-source code for the methodology described. |
| Open Datasets | No | The paper uses a theoretical model for data generation: 'We assume that the real data yk Rn, k = 0, 1, . . . are drawn according to the following generative model: yk = G(ck,ak; U, ηT) def = Uck + ηTak'. It does not use a publicly available, real-world dataset. The numerical verification uses simulated data according to these models. |
| Dataset Splits | No | The paper performs numerical simulations to verify theoretical predictions but does not specify dataset splits such as train/validation/test percentages or counts, as it uses simulated data generated according to theoretical models. |
| Hardware Specification | No | The paper mentions 'numerical simulations' but does not provide any specific hardware details such as GPU/CPU models, memory, or cloud computing instances used for running these simulations. |
| Software Dependencies | No | The paper does not specify any software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions or other libraries) used for the numerical simulations. |
| Experiment Setup | Yes | The number of features is d = 2, and ck and eck are both Gaussian with zero mean and covariance diag([5, 3]). The dimension is n = 5, 000, and the learning rates of the generator and discriminator are eτ = 0.04 and τ = 0.2 respectively. After testing different noise strength ηT = ηG = 2, 1, 4... |