Seminar with PhD Student Adji Bousso Dieng: Prescribed Generative Adversarial Networks
Seminar | October 18 | 11 a.m.-12 p.m. | Sutardja Dai Hall, Room 250
Adji Bousso Dieng, Columbia University
GANs are a powerful approach to unsupervised learning. They have achieved state-of-the-art performance in the image domain. However, GANs are limited in two ways. They often learn distributions with low support---a phenomenon known as mode collapse---and they do not guarantee the existence of a probability density, which makes evaluation with predictive log-likelihood impossible. In this talk, I will present the prescribed GAN (PresGAN) which addresses these shortcomings. PresGANs add noise to the output of a density network and optimize an entropy-regularized adversarial loss. The added noise renders tractable approximations of the predictive log-likelihood and stabilizes the training procedure. The entropy regularizer encourages PresGANs to capture all the modes of the data distribution. Fitting PresGANs involves computing the intractable gradients of the entropy regularization term; PresGANs sidestep this intractability using unbiased stochastic estimates. I will then showcase the performance of PresGANs on several datasets in terms of how they mitigate mode collapse, generate samples with high perceptual quality, and achieve competitive predictive performance as measured by log-likelihood when compared to VAEs in a controlled experiment.
Bio: Adji is a Ph.D student in the department of Statistics at Columbia University where she is jointly being advised by David Blei and John Paisley. Her doctoral work is about deep generative models. More specifically, she designs algorithms for fitting deep generative models and combines probabilistic modeling and deep learning to embed structure into deep generative models.
Prior to joining Columbia she worked as a Junior Professional Associate at the World Bank. She did her undergraduate training in France where she attended Lycee Henri IV and Telecom ParisTech--France's Grandes Ecoles system.
CA, email@example.com, 4155196086