If you have seen the last video on GANs, it may have created an illusion that training GANs is super easy. But when you sit back and train them, they can be quite a headache. Stability of Generator is extremely fragile in most of the scenarios. So today, we are going to present a well-written paper by Martin Arjovsky, Soumith Chintala and L´eon Bottou from Facebook AI Research Group.

I am also impressed by blogs written by Vincent Herrmann and Alex Irpan.


Video Covers :

Why do we need Earth Mover's Distance?

How does WGANs or wasserstein gans gets rid of Mode Collapse? (ofcourse clipping weights is brutal)

Why is training WGANs more stable?

What is kantorovich Rubinstien duality?

Cool visualizations to understand difficulties in estimation of a probability distribution.

References

1. https://vincentherrmann.github.io/blo...

2. https://www.alexirpan.com/2017/02/22/...

3. https://lilianweng.github.io/lil-log/...

4. Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., & Courville, A. C. (2017). Improved training of wasserstein gans. In Advances in Neural Information Processing Systems (pp. 5767-5777).

5. https://arxiv.org/abs/1701.07875

6. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., & Chen, X. (2016). Improved techniques for training gans. In Advances in Neural Information Processing Systems (pp. 2234-2242).

Please consider supporting the channel on Patreon page :
  / crazymuse  

Thanks to my current patrons for supporting the work :
1. Parth Parikh
2. Laher .D
3. Sean Marrett