In Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy
will defend his proposal
Self-Competitional Generative Adversarial Neural Networks
AbstractIn this research, we propose different types of new data-driven and algorithmic approaches to train the Generative Adversarial Nets (GANs). In the data-driven approach, we introduce a new concept called virtual big data. We prove that the virtual big data can provide the GANs sufﬁcient training data to generate efﬁcient augmented data with less mode collapse and vanishing generator gradients problems. We show that the curse of dimensionality, which is considered as a negative factor in machine learning, can play a positive role to solve vanishing generator gradients via making the discriminator less perfect. First, we transform the training data from n dimensional space into m dimensional space where, m = c∗n and c is the concatenation factor. To do so, c different training instances are selected and concatenated to each other to form a c∗n dimensional instance. Increasing the dimension of training data from n to c∗n is key to increase the number of training instances from N to C(N,c). Transformed training data are called virtual big data since they differ with original training instances in terms of size and dimension. Our experiments show that V-GAN, a GAN trained by virtual big data, can outperform standard GANs when it comes to dealing with extremely scarce training data. Furthermore, V-GAN can outperform traditional oversampling techniques in terms of precision, F1 score and Area Under Curve (AUC) score. In an algorithmic approach, we propose Self-Competitional GANs (SCOM-GANs) in which the GAN competes with itself to generate better results compared to the previous epoch. To create the SCOM-GAN, we add a supervised unit (softmax) to the discriminator as an auxiliary to stabilize the GAN training without increasing the training data size. To train the auxiliary classifier we need pseudo-labels which are provided by keep tracking of the fake data in c consecutive epochs and labelling them based on (epoch number % c). This proposed process, which is called dynamic labelling, lets the discriminator start training with limited labels. Afterwards, more training data with different labels are added during the next epochs. The role of auxiliary classifier is to distinguish the new fake data generated in the current epoch from the old ones generated in previous epochs. This way, we provide two different types of supervisory signals for the generator: real/fake signals and old/new signals. Therefore, if the generator starts generating the same instances in consecutive epochs, then it would be punished with higher loss values thanks to the auxiliary classifier. The discriminator and classifier are the same (shared parameters), except for the last layer. Adding the auxiliary classifier to distinguish old fakes and new fakes has two potential benefits. First, it can decrease the vanishing generator gradient since the SCOM discriminator is a multi-task neural network with hard parameter sharing which negative transfers caused by unrelated tasks can make the discriminator less perfect. Second, SCOM-GAN can potentially avoid mode collapse since the generator is punished if it starts generating redundant data in consecutive epochs.
Date: Wednesday, May 20, 2020
Time: 2:00 - 3:00 PM
Place: Online Presentation - Zoom Meeting
To register for attending this session, you need to send a request to email@example.com via academic email at least 6 hours before starting the online meeting.
Advisor: Dr. Larry Shi
Faculty, students, and the general public are invited.