AI-chitect

AI-chitect – 人工智能建筑师

Future Architecture Lab – 未来建筑实验室

Artificial intelligence technology nowadays such as Generative Adversarial Neural Networks (GAN) is able to achieve human’s creativity to generate anything based on mechanism that set by human. The model mechanism consists of one “generator” that synthesizes “fake” samples from scratch, and a “discriminator” that evaluates the generated fake samples if they are “real (similar to a real sample)” or “fake (does not fit to any real sample)”. Generator and discriminator learn from each other, hence higher discrimination level ensures higher quality of generation. Such mechanism makes machine capable to synthesis images that are almost similar to what you can found in real world.

We need real sample dataset to train GAN model. Human needs to control the quality of samples as to ensure the generated output could be reflect on those real sample in some extent. This exhibition’s output is a new kind of “architectural representation”.Trained model needs new input as trigger to carry out generation. Input here normally is digital image, in this exhibition input will be a roughly sketch silhouette that drawn by human on a sketchpad. Based on input, machine could start its generation without human intervention. Hence, it is only machine itself to be participated in the most “creative” generation part, and human will not know what the outcome will be.

Uncertainty exists in generation phases; the unpredictable generated images might inspire one’s imagination or to provoke extensive thinking. Human do not limit themselves in interpretation from these generated images. Such process could analogous to sketching, where one could keep developing his or her design from the sketches on paper. Sketching involves communication between eyes, hand and brain, allowing design idea develops either from narrow to broad, or inversely. The installation is taken part in diverging one’s thinking. Hence, this indicates that, by integration of GAN, machine could assist human to broaden creative thinking. Besides, this interactive human-machine installation might also could “reconstruct” the design method or workflow that architects have been practiced for several centuries.

当今人工智能技术，如生成对抗性神经网络（Generative Adversarial Neural Networks, GAN）已能实现人的创造性，即根据人为预设的机制生成任何事物。该模型机制由一个“生成器“和一个”判别器”所组成，前者从零生成“伪”样本，后者则评估所生成的伪样本是否真实（与某真实样本相似）或虚假（不匹配任何真实样本）。生成器和判别器相互学习，判别水平越高，生成质量越高。这种机制使机器能够生成与现实世界几乎相似的图像。

我们需要真实的样本数据集训练GAN模型。人类控制样本的质量，以确保生成的输出在一定程度上能够反映真实的样本。本次展览中的输出为其中一种新型的“建筑表现图”。训练好的模型需要新的输入作为触发器进行生成。而输入则是由人类在速写板上绘制的草图轮廓。机器可以基于输入，在没有人工干预的情况下开始生成新的图示。因此只有机器本身参与到具有“创造性”的生成过程中，而人类对此无法预测。

不确定性存在于生成阶段，这些无法预测的生成图像可以激发人的想象力或拓展思考。人类并不局限自己对这些生成图像上的诠释。这样的过程类似于素描，我们可以通过纸上的素描发展设计。素描过程涉及到了眼睛、手和大脑之间的交流，将设计思维从一点拓展到多种可能，抑或是相反。该装置则是参与到发散思维的过程，因此这似乎体现GAN可以帮助人类拓展创造性的思维。除此之外，该交互式装置还或许可以“重构”建筑师已经实践了数世纪的设计方法或工作流程。

Credits:

Wanyu He 何宛余, Xiaodi Yang 杨小荻, Jackie Yong Leong Shong 杨良崧, Kan Liu 刘勘.