À la Une

Open research challenges for GAN training

Speaker: Dr. Ian Goodfellow, Google Brain

IG.jpg

Title: Open research challenges for GAN training

Abstract:

Recent work has shown that simultaneous gradient descent on both players' costs converges to the Nash equilibrium of the GAN game. Unfortunately, other recent work has shown that this convergence may be very slow. Future research work should aim to either reparameterize the game to make convergence faster, or design new equilibrium-finding algorithms that are able to converge rapidly in this structure of game.

Short bio:

Dr. Ian Goodfellow is the inventor of Generative Adversarial Networks, which according to Yan LeCun is the "most interesting idea in machine learning in the last 10 years". He is the main author of the deep learning book. He did his PhD in the University of Montreal with Yoshua Bengio and Aaron Courville in machine learning. He then worked in Google Brain and OpenAI. He is currently with Google.

Date: Friday 27th October 2017, 9.00 am

Location: Battelle building B - room 4.11 (4th floor) (map)