Is Discriminator a Good Feature Extractor?

• Main Authors: Mao, Xin, Su, Zhaoyu, Tan, Pin Siang, Chow, Jun Kang, Wang, Yu-Hsing
• Format: Journal Article
• Language: English
• Published: 02-12-2019
• Subjects: Computer Science - Learning; Statistics - Machine Learning
• DOI: 10.48550/arxiv.1912.00789

The discriminator from generative adversarial nets (GAN) has been used by
researchers as a feature extractor in transfer learning and appeared worked
well. However, there are also studies that believe this is the wrong research
direction because intuitively the task of the discriminator focuses on
separating the real samples from the generated ones, making features extracted
in this way useless for most of the downstream tasks. To avoid this dilemma, we
first conducted a thorough theoretical analysis of the relationship between the
discriminator task and the features extracted. We found that the connection
between the task of the discriminator and the feature is not as strong as was
thought, for that the main factor restricting the feature learned by the
discriminator is not the task, but is the need to prevent the entire GAN model
from mode collapse during the training. From this perspective and combined with
further analyses, we found that to avoid mode collapse, the features extracted
by the discriminator are not guided to be different for the real samples, but
divergence without noise is indeed allowed and occupies a large proportion of
the feature space. This makes the features more robust and helps answer the
question as to why the discriminator can succeed as a feature extractor in
related research. Consequently, to expose the essence of the discriminator
extractor as different from other extractors, we analyze the counterpart of the
discriminator extractor, the classifier extractor that assigns the target
samples to different categories. We found the performance of the discriminator
extractor may be inferior to the classifier based extractor when the source
classification task is similar to the target task, which is the common case,
but the ability to avoid noise prevents the discriminator from being replaced
by the classifier.