Oral
GDPP: Learning Diverse Generations using Determinantal Point Processes
Mohamed Elfeki · Camille Couprie · Morgane Riviere · Mohamed Elhoseiny

Tue Jun 11th 04:30 -- 04:35 PM @ Seaside Ballroom

Generative models have proven to be an outstanding tool for representing high-dimensional probability distributions and generating realistic looking images. An essential characteristic of generative models is their ability to produce multi-modal outputs. However, while training, they are often susceptible to mode collapse, that is models are limited in mapping the input noise to only a few modes of the true data distribution. In this paper, we draw inspiration from Determinantal Point Process (DPP) to propose an unsupervised penalty loss that alleviates mode collapse while producing higher quality samples. DPP is an elegant probabilistic measure used to model negative correlations within a subset and hence quantify its diversity. We use DPP kernel to model the diversity in real data as well as in synthetic data. Then, we devise an objective term that encourages the generator to synthesize data with a similar diversity to real data. In contrast to previous state-of-the-art generative models that tend to use additional trainable parameters or complex training paradigms, our method does not change the original training scheme. Embedded in an adversarial training and variational autoencoder, our Generative DPP approach shows a consistent resistance to mode-collapse on a wide-variety of synthetic data and natural image datasets including MNIST, CIFAR10, and CelebA, while outperforming state-of-the-art methods for data-efficiency, convergence-time, and generation quality whereas being 5.8x faster than its closest competitor. Our code, attached to the submission, will be made publicly available.

Author Information

Mohamed Elfeki (University of Central Florida)
Camille Couprie (FAIR)
Morgane Riviere (Facebook Artificial Intelligence Research)
Mohamed Elhoseiny (KAUST and Baidu SVAIL)

Dr. Mohamed Elhoseiny is Assistant Professor of Computer Science at the Visual Computing Center at KAUST (King Abdullah University of Science and Technology) and an AI Research consultant at Baidu Research at Silicon Valley AI Lab (SVAIL). He received his PhD from Rutgers university under Prof. Ahmed Elgammal in October 2016 then spent more than two years at Facebook AI Research(FAIR) until January 2019 as a Postdoc Researcher. His primary research interests are in computer vision and especially about learning about the unseen or the least unseen by recognition (e.g., zero-shot learning) or by generation (creative art and fashion generation). Under the umbrella of how AI may benefit biodiversity, Dr. Elhoseiny's 6-years long development of the zero-shot task on major vision conferences was featured in the United Nations biodiversity conference in November 2018 (~10,000 audience from >192 countries). His creative AI research projects were recognized at the ECCV18 workshop on Fashion and Art with the best paper award, media coverage at the New Scientist Magazine and MIT Tech review (2017, 2018), 20 min speech at the Facebook F8 conference (2018), and coverage at HBO Silicon Valley TV Series (2018), and the official FAIR video

Related Events (a corresponding poster, oral, or spotlight)