课程简介
GAN生成式对抗网络
目标收益
培训对象
对深度学习算法原理和应用感兴趣，具有一定编程(Python)和数学基础(线性代数、微积分、概率论)的技术人员。
对深度学习模型，特别是生成式模型（有一定了解为佳。
课程大纲
| 1. GAN 入门 - Generative Models |
- Latent Factors - Generative Models |
| 2. GAN 原理 |
- Discrimination and Generator - Training GAN - Distances in GAN: from KL-Divergence to JS-Divergence and others - Problems with GAN - CGAN and DCGAN |
| 3. f-GAN 模型 - GAN 模型的同一框架 |
- Fenchel Conjugate - f-Divergence - Training: double loop vs single loop - 多种 divergence 函数 |
| 4. Wasserstein GAN - WGAN |
- Problem with JS-Divergence - Mode Collapse - Earth-Mover Distance - WGAN - EBGAN: Energy-Based GAN |
| 5. InfoGAN - 可解释表示的 GAN |
- 潜因子与表象的互信息 - 现有 GAN 和 Domain 之间的矛盾 - 用无监督学习发现可解释的潜因子 - 带互信息正则项的 loss 函数 - 实现：用变分法进行训练 - 效果 |
| 6. GAN for NLP |
- Improving Sequence Generation with GAN - Conditional Sequence Generation - Unsupervised Conditional Sequence Generation |
| 7. GAN for CV |
- GAN + Autoencoder: Photo Editing - Image Super Resolution - Image Completion |
|
1. GAN 入门 - Generative Models - Latent Factors - Generative Models |
|
2. GAN 原理 - Discrimination and Generator - Training GAN - Distances in GAN: from KL-Divergence to JS-Divergence and others - Problems with GAN - CGAN and DCGAN |
|
3. f-GAN 模型 - GAN 模型的同一框架 - Fenchel Conjugate - f-Divergence - Training: double loop vs single loop - 多种 divergence 函数 |
|
4. Wasserstein GAN - WGAN - Problem with JS-Divergence - Mode Collapse - Earth-Mover Distance - WGAN - EBGAN: Energy-Based GAN |
|
5. InfoGAN - 可解释表示的 GAN - 潜因子与表象的互信息 - 现有 GAN 和 Domain 之间的矛盾 - 用无监督学习发现可解释的潜因子 - 带互信息正则项的 loss 函数 - 实现：用变分法进行训练 - 效果 |
|
6. GAN for NLP - Improving Sequence Generation with GAN - Conditional Sequence Generation - Unsupervised Conditional Sequence Generation |
|
7. GAN for CV - GAN + Autoencoder: Photo Editing - Image Super Resolution - Image Completion |
近期公开课推荐
