Bridging Maximum Likelihood and Adversarial Learning via α-Divergence
Authors: Miaoyun Zhao, Yulai Cong, Shuyang Dai, Lawrence Carin6901-6908
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the proposed α-Bridge from three perspectives. First we show that the α-Bridge, dynamically transferring advantages from ML to adversarial learning, exhibits a more stable training with improved robustness to hyperparameters (this is expected because of the aforementioned discussions of control variants interpretation and connections to prior GAN regularization methods). We then show that the α-Bridge is capable of smoothly transferring the information learned during ML learning to adversarial learning, circumventing the forgetting issue shown in Figure 1. Finally we highlight the versatility of the α-Bridge, by showing its capability in transplanting the variational approximation within ML learning into an inference arm for adversarial learning. See Appendix G for the detailed experimental settings and the corresponding analysis/discussions. |
| Researcher Affiliation | Academia | Miaoyun Zhao, Yulai Cong,* Shuyang Dai, Lawrence Carin Department of Electrical and Computer Engineering, Duke University |
| Pseudocode | Yes | Algorithm 1 α-Bridge (from forward to reverse KL) |
| Open Source Code | No | The paper does not contain any explicit statement about releasing source code or a link to a code repository. |
| Open Datasets | Yes | The 25-Gaussians example from (Tao et al. 2018) is adopted, where the data are generated from a 2D Gaussian mixture model with 25 components, as shown in Figure 3. ... To address the concern of how α-Bridge performs on real datasets, we conduct another experiment on CIFAR10 (Krizhevsky and Hinton 2009) and observe an improved performance... we run α-Bridge on the MNIST (Le Cun et al. 1998) and Celeb A (Liu et al. 2015) datasets... |
| Dataset Splits | No | The paper mentions using specific datasets for experiments but does not provide details on training, validation, or test dataset splits (e.g., percentages, sample counts, or cross-validation setup). |
| Hardware Specification | Yes | The Titan Xp GPU used was donated by the NVIDIA Corporation. |
| Software Dependencies | No | The paper mentions using TensorFlow and PyTorch, but it does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | Adam (Kingma and Ba 2014) hyperparameter β1 = 0.1 / β1 = 0.5. |