| Description |
-
This thesis concerns
the current state of Generative Adversarial Networks (GANs) in the field of Machine
Learning. The first part of the work combines the BayesGAN and the BiGAN models to
create the Bayesian Bidirectional GAN (BayesBiGAN). The BayesBiGAN uses the BayesGAN's
MCMC training method to sample from posteriors over the generator, discriminator, and
the BiGAN encoder. This model ... read morecaptures a rich underlying feature representation and
encodes signals to a lower dimensional vectors. The second part of the work addresses
the disconnect between GAN research and product development. GAN research is far less
accessible than research in other machine learning fields due to a lack of developer
tools. This drastically limits the speed and breadth of applications that can be made.
This work begins to address the disconnect by creating an API that offers cloud access
to CycleGAN for style transfer to convert photos into painting renderings in the style
of various famous painters. The API allows developers to easily integrate the CycleGAN
into their own work. An educational demo website demonstrates how the API can be used to
engage a non-technical audience.
Thesis
(M.S.)--Tufts University, 2019.
Submitted to the
Dept. of Computer Science.
Advisor: Liping
Liu.
Committee: Liping Liu, Samuel Guyer, and
Xiaozhe Hu.
Keyword: Computer
science.read less
|
|---|
This object is in collection
