Dynamic Neural Architecture Warping for Time Series Recognition

2022 Fiscal Year Annual Research Report

• Project/Area Number: 21K17808
• Research Institution: Kyushu University
• Principal Investigator: IWANA BRIAN・KENJI 九州大学, システム情報科学研究院, 准教授 (90852723)
• Project Period (FY): 2021-04-01 – 2023-03-31
• Keywords: Time series / Temporal neural network / Dynamic programming

Outline of Annual Research Achievements

This research aims to develop dynamic neural networks. Dynamic neural networks are neural networks that can adapt based on the input. A number of papers were successfully published. This includes three international conference papers, from ICASSP, ICPR, and ACMMM, and two Pattern Recognition journal papers. All of these papers are highly respected publications. There are a number of domestic publications as well.

Significant progress was done in dynamic temporal neural networks. Besides proposing novel models, new applications with real data were addressed. For example, a paper published at NLP 2023 used a new corpus that was created with a collaborator.

Research Products

• [Journal Article] FETNet: Feature erasing and transferring network for scene text removal (2023)
  • Author(s): Guangtao Lyu, Kun Liu, Anna Zhu, Seiichi Uchida, Brian Kenji Iwana
  • Journal Title: Pattern Recognition Volume: 140 Pages: 109531～109531
  • DOI: 10.1016/j.patcog.2023.109531
  • Peer Reviewed / Int'l Joint Research
• [Journal Article] Deep attentive time warping (2023)
  • Author(s): Shinnosuke Matsuo, Xiaomeng Wu, Gantugs Atarsaikhan, Akisato Kimura, Kunio Kashino, Brian Kenji Iwana, Seiichi Uchida
  • Journal Title: Pattern Recognition Volume: 136 Pages: 109201～109201
  • DOI: 10.1016/j.patcog.2022.109201
  • Peer Reviewed / Int'l Joint Research
• [Presentation] Classification of Polysemous and Homograph Word Usages using Semi-Supervised Learning (2023)
  • Author(s): Sangjun Han, Brian Kenji Iwana, Satoru Uchida
  • Organizer: Annual Conference of the Association for Natural Language Processing (NLP)
• [Presentation] Text Style Transfer based on Multi-factor Disentanglement and Mixture (2022)
  • Author(s): Anna Zhu, Zhanhui Yin, Brian Kenji Iwana, Xinyu Zhou, Shengwu Xiong
  • Organizer: ACM Multimedia (ACMMM)
  • Int'l Joint Research
• [Presentation] Dynamic Data Augmentation with Gating Networks for Time Series Recognition (2022)
  • Author(s): Daisuke Oba, Brian Kenji Iwana, Shinnosuke Matsuo
  • Organizer: International Conference on Pattern Recognition (ICPR)
  • Int'l Joint Research
• [Presentation] On Mini-Batch Training with Varying Length Time Series (2022)
  • Author(s): Brian Kenji Iwana
  • Organizer: International Conference on Acoustics, Speech, and Signal Processing (ICASSP)
  • Int'l Joint Research
• [Presentation] Word difficulty-level analysis using machine learning (2022)
  • Author(s): Sangjun Han, Brian Kenji Iwana
  • Organizer: 情報関係学会九州支部連合大会
• [Presentation] Making time series neural networks robust to attacks (2022)
  • Author(s): Yoh Yamashita, Brian Kenji Iwana
  • Organizer: 情報関係学会九州支部連合大会