Enhancing Sample Diversity in Deep Convolutional GANs Using a Bayesian Framework
Abstract: Generative Adversarial Networks (GANs) are proficient at generating synthetic data; how-
ever, they face the ongoing issue of mode collapse. This problem arises when the generator
focuses on producing a limited set of data patterns that trick the discriminator during the
optimization process, requiring novel solutions. In our endeavor to address mode collapse in
Deep Convolutional Generative Adversarial Networks (DCGAN) and foster greater sample
diversity, we introduce a Bayesian framework applied to DCGAN, referred to as Bayesian
DCGAN. This framework makes three key contributions: (i) the integration of a weight
distribution within the network, achieved through the application of the Bayes by Backprop
method; (ii) employing a mean-field variational inference approach to approximate the pos-
terior distributions of weights; and (iii) putting forth a mathematical approach to quantify
the diversity present in the samples generated by Bayesian DCGAN, contrasting it with the
output of conventional DCGAN. Our experimental results showcase that Bayesian DCGAN
generates more diverse samples compared to its conventional counterpart, thereby signifi-
cantly reducing uncertainty in neural networks. This enhancement in diversity is pivotal for
creating robust and adaptable models, particularly in scenarios where a broader spectrum
of data representations is essential for effective learning and generalization.
Submission Length: Long submission (more than 12 pages of main content)
Assigned Action Editor: ~Ruoyu_Sun1
Submission Number: 3005
Loading