A user-interface to explore the latent space of Generative Adversarial Networks
Keywords
Loading...
Authors
Issue Date
2021-01-29
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
Generative Adversarial Networks (GANs) have recently revolutionized arti cial image synthesis,
which also led to a promising application of GANs to create large datasets for machine learning,
neuroscience, and beyond. Furthermore, these breakthroughs led to discovering various techniques
to understand how GANs learn the latent representation. One of them is to identify linear subspaces
within the latent space. A subspace represents a set of characteristics from which directional vectors
are extractable. These vectors enable the generation of customized datasets. However, the latent
space is a complex entity where a subspace may warp, which requires ne-tuning of every vector
application to achieve naturally-looking images. Thus, at the moment, no algorithmic method could
be directly applied to generate such a custom dataset. Naturally, the problem poses a question of
how much should each vector be scaled to enhance a feature and still make images look credible? In
this work, we propose a user interface to simplify latent space navigation. Additionally, we o er a
solution where a human is included in the loop to recognize the most natural-looking image using a
slider that navigates the latent space along a vector direction. This method allows pinpointing the
location in the latent subspace where the attribute looks most credible.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen