Deep Learningを用いたデザインAIの作成と検証

街並みと建築物外観の画像生成を対象に

抄録

 In recent years, there has been a growing interest in the application of Deep Learning to architecture and urbanism. This research is focused on content generation AI using Deep Learning. Despite claims that replacing creativity-related work with machines is difficult, the use of generative adversarial networks (GANs) is becoming more popular in various fields. The objective of this research is to develop an AI-supported design or a co-creation between humans and AI through the application of GANs. The primary goal of this work could be interpreted as repurposing existing concepts to create new designs through the combination of multiple design sources. Therefore, the purpose of this research is the creation of AI that emulate and support the design process.

 

 This research examines two types of AI through a two-stage process; the first is an AI that reproduces design, and the second is an AI that generates design. The first type of AI reproduces designs from different sources and includes an analysis of whether the design can be expressed mathematically. This analysis is a prerequisite for the creation of the second type of AI that generates new designs by combining information from multiple sources. In other words, the second type of AI views designs mathematically, and the possibility of expressing designs mathematically (using the first type of AI) is examined to ensure that such a function is feasible and in line with user intention. Here, a mathematical expression refers to a 100-dimensional vector and an already-learned deep neural network.

 

 The AI that reproduces design was applied to famous cityscapes (Kyoto and Edinburgh) and the façades of famous buildings (three works by Le Corbusier). The designs were reproduced as images and used for subject experiments to confirm that the intended impressions (oriental and occidental) and the designs of each type were successfully reproduced.

 

 For the AI that generates design, a new design was generated from calculations of different combinations (three pairs and one trio) of the façades of three works by Le Corbusier (church of Saint-Pierre, Notre Dame du Haut, and Villa Savoye). This design was subsequently used for text mining Bayesian-estimation-based subject experiments to confirm that the characteristics of the design sources were successfully inherited.

 

 To the best of our knowledge, these are new types of AI. Further, we believe that these achievements may facilitate better dissemination of design through fast generation of a large number of images (design patterns) that constitute new types of designs. This achievement may also help expand the concept of human design thinking by suggesting designs that can be permuted using AI but otherwise inconceivable for human designers. Ultimately, this can help in the creation of a new design environment, namely “co-creation between humans and AI,” wherein the designers choose the sources and the AI generates a number of design choices for the final design.