Abstract: Generative Adversarial Net (GAN) is a powerful machinelearning model, and becomes extremely successful recently. The generator andthe discriminator in a GAN model competes each other and reaches the Nashequilibrium. GAN can generate samples automatically, therefore reduce therequirements for large amount of training data. It can also model distributionsfrom data samples. In spite of its popularity, GAN model lacks theoreticfoundation.

In this talk, we give a geometric interpretation to optimal mass transportation theory, and applied for the GAN model. We try to answer the following fundamental questions:

1. Does a GAN model learn a function, a mapping or a probability distribution ? Is the solution unique or infinite ? What is the dimension of the solution space ? What is the structure of the solution space?

2. Does a GAN model really learn or just memorize ? 

3. Is the competition between the generator and discriminator really necessary ? Can we simplify the neural networks and avoid the competition ?

4. Why sometimes a ML model can be fooled easily ?

5. Can we replace the black-box in the GAN model by a transparent model ?

Presenter: Professor David Gu. 

Dr. Gu is a tenured professor in Computer Science Department and Applied Mathematics Department of State University of New York at Stony brook. Dr.Gu is one of the major founders of an emerging interdisciplinary field:computational conformal geometry, which applies modern geometry in engineeringand medicine fields.  He won Morningside Gold medal in applied mathematicsfor this work in 2013. Dr. Gu has published about 300 articles in toplevel journals and conferences in graphics, vision, visualization, medicalimaging and networking fields; four monographs in mathematics and computerscience. He has obtained several international patents, some of them have beenlicensed to Simens and GE.

CSE 600 students will receive credit for attending.