[3E4-OS-12b] 画像とAI（MIRU2019プレビュー）(2)

[3E4-OS-12b-03] A Generative Framework for Creative Data Based on the Generative Adversarial Networks

〇Riku Fujimoto¹, Takato Horii¹, Tatsuya Aoki¹, Takayuki Nagai^1,2 (1. The University of Electro-Commnications, 2. Osaka University)

Keywords:Generative adversarial nets, Creativity, Value

In this research, we propose a framework to generate creative data simulating the creation process. This framework generates new and valuable high dimensional data.The characteristics of this framework are two points, a mixed generator and self-generated data learning. The mixed generator makes it possible to generate new data by loss function of regularization by Feature matching and entropy. In self-generated data learning, expressive ability to generate higher value data is acquired by using highly valued generation data as learning data.The framework for generating new and valuable data by combining these two methods is called "Deep Creative Model(DCM)".In the experiments, MNIST was used as learning data, learning a framework to set alphabet images as valuable images. As learning progresses, it was possible to gradually generate images close to the shape of the alphabet, and it was confirmed that it is possible to generate creative data with DCM.