Why do generative models (not) work? Online meetup

Philipp Kofman, Deep Learning researcher. 

This year Philipp has finished master’s program at Ukraine Catholic University. During studying at the university, he took part in ACM ICPC — International programming competition and archived great results on semi-final several times. This was facilitated by studying in one another place — Yandex school of data analysis. His passion for algorithms and complex tasks combined with the desire to solve business problems and led him to work in the field of data analysis. Philipp has worked for more than 3 years in data analysis in companies such as MMI and TrueConf. His professional interests include deep generative models, before that, he worked on time series prediction problems. 

In this lecture, he would like to share basic knowledge about simple generative models, show how they work and develop your intuition. Modern generative models can make a miracle… Dozens of research papers describe magic methods for animals, cool images, etc generation. However, in practice, we face a lot of problems and pain during model training.

Models don’t have a good generation property due to: bad dataset, small capacity of neural nets, too simple or too complex distribution of generated data and a lot of other conditions. In this lecture you will find out some tricks for improving models, dive deeper into the understanding of complex generation pipeline, better understand problems and how we can solve it (or not…).

If you are not sure, try to solve the 4th case from this site https://poloclub.github.io/ganlab/