[HKML] Hong Kong Machine Learning Meetup Season 1 Episode 12 (Season Finale)
- Wednesday, July 17, 2019 from 7:00 PM to 9:00 PM
- Amazon AWS Hong Kong, Tower 535, aws pop-up loft 26/F, Causeway Bay, Hong Kong
Generative Adversarial Network (GAN) is notorious for being hard to train. Wasserstein GAN with Gradient Penalty, as one of its variants, comes handy into the rescue. The presentation discussed a number of major problems one commonly faces when training a GAN: non-smooth cost function, vanishing gradients, mode collapse, lack of indicative metrics for training performance (on this issue, I learned about the Geometry Score: A Method For Comparing Generative Adversarial Networks, at ICML 2018, based on topological features; I did not have time yet to experiment with it to see if it really helps in practice). After the long and precise introduction, Alex explained how WGAN-GP can tackle some of these problems. A few applications of WGAN-GP were highlighted.
His slides are there.
Alex is looking for collaboration on fun and non-commercial projects he has in mind. Reach out to him for further information if you are interested.
Eric Greene - Forecasting Time Series with AWS
Slides will come soon.
Gautier Marti - Takeaways from ICML 2019, Long Beach, California
I presented very briefly the content of the following slides. I tried to identify some niche trends emerging in the ML community. No focus on deep (reinforcement) learning at all, which was a big part of ICML though.