Event

Mathematics & Statistics Graduate Student Seminar: how optimal transport and neural networks can generate photos of anime characters

Friday, November 23, 2018 13:00to14:00
Burnside Hall Room 1025, 805 rue Sherbrooke Ouest, Montreal, QC, H3A 0B9, CA

Hi, everyone,

 

The Mathematics & Statistics Graduate Student Seminar will convene at 1:00pm this Friday, November 23, in the Main Lounge (Burnside 1025). As usual, there will be pizza.

 

This week, Aram will talk to us about how optimal transport and neural networks can generate photos of anime characters:

 

This is another talk about how mathematical tools can be used in Deep Learning. We will begin with a brief introduction on Optimal Transport and generative adversarial networks, their connection and how they can be applied to anime characters.

Optimal Transport (OT) is a concept that has been around since Gaspard Monge published ''Sur la théorie des déblais et des remblais" in the late 1700s. Roughly, this translates to taking a pile of sand and moving it somewhere else (it's not actually about sand but probability measures).

How can this possibly tie into deep neural networks? Generative adversarial models (GANs) are an application of deep neural networks; as the name indicates, the goal is to generate a distribution that resembles that of "real" data. GAns are in their infancy, as they were introduced in 2014. GAns require a "loss function" to train the networks to do a better job at generating the data. Researches last winter introduced OT into training GANs, giving some better convergence properties!

 

See you all there!

 

All graduate students are invited. As with all talks in the graduate student seminar, this talk will be accessible to all graduate students in math and stats.This seminar was made possible by funding from the McGill Mathematics and Statistics Department and PGSS.

Follow us on

Back to top