Impact of Duplicating Small Training Data on GANs

Yuki Eizuka, Kazuo Hara, Ikumi Suzuki

2021

Abstract

Emoticons such as (_̂)̂ are face-shaped symbol sequences that are used to express emotions in text. However, the number of emoticons is miniscule. To increase the number of emoticons, we created emoticons using SeqGANs, which are generative adversarial networks for generating sequences. However, the small number of emoticons means that few emoticons can be used as training data for SeqGANs. This is concerning because as SeqGANs underfit small training data, generating emoticons using SeqGANs is difficult. To address this problem, we duplicate the training data. We observed that emoticons can be generated when the duplication magnification is of an appropriate value. However, as a trade-off, it was also observed that SeqGANs overfit the training data, i.e., they produce emoticons that are exactly the same as the training data.

Paper Citation

in Harvard Style

Eizuka Y., Hara K. and Suzuki I. (2021). Impact of Duplicating Small Training Data on GANs. In Proceedings of the 10th International Conference on Data Science, Technology and Applications - Volume 1: DATA, ISBN 978-989-758-521-0, pages 308-315. DOI: 10.5220/0010583403080315

in Bibtex Style

@conference{data21,
author={Yuki Eizuka and Kazuo Hara and Ikumi Suzuki},
title={Impact of Duplicating Small Training Data on GANs},
booktitle={Proceedings of the 10th International Conference on Data Science, Technology and Applications - Volume 1: DATA,},
year={2021},
pages={308-315},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0010583403080315},
isbn={978-989-758-521-0},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 10th International Conference on Data Science, Technology and Applications - Volume 1: DATA,
TI - Impact of Duplicating Small Training Data on GANs
SN - 978-989-758-521-0
AU - Eizuka Y.
AU - Hara K.
AU - Suzuki I.
PY - 2021
SP - 308
EP - 315
DO - 10.5220/0010583403080315