The presentation by Naouhei Taniguchi, a first-year master’s student in our laboratory, was selected for the Student Incentive Award at the 2019 National Conference of the Japanese Society for Artificial Intelligence.
Title] Generative Query Network as Meta-Learning
Summary]
Generative Query Network (GQN) is an innovative deep generative model that enables rendering of observed images from unknown viewpoints and has attracted attention as a new 3D modeling method. However, GQN is known to have some issues, such as a huge training cost and unstable training due to its sensitivity to hyperparameters. In addition, the lack of validation of GQN as a probabilistic model has hindered developmental research due to the low interpretability of the model architecture. To address these issues, we formulate a probabilistic model of GQN using a meta-learning framework and propose a method to improve the cost and instability of learning based on it. Evaluation experiments were conducted on the Shepard Metzler dataset to verify the effectiveness of the proposed method.
Author] Shohei Taniguchi, Yusuke Iwasawa, Yutaka Matsuo
[paper link] https://www.jstage.jst.go.jp/article/pjsai/JSAI2019/0/JSAI2019_2Q5J203/_article/-char/ja
