2020.07.20 Monday

Deep Learning Architect: AI Classification for Architectural Design


Deep convolutional neural network

We utilized NASNet, a novel program that achieves state-of-the-art accuracy while halving the computational cost of the best reported results. NASNet is composed of two types of layers: a normal layer and a reduction layer, both designed by auto machine learning, which is the automated process for constructing models.


It presents an example of Grad-CAM applied to an image taken from Alva Aalto’s work. The left picture is the original, and the right pictures are the results of Grad-CAM for each architect. The red color indicates the location where the machine’s eye focuses in order to identify the similarity of the design with the original picture. For example, the computer vision reacted to Miralles and Lloyd due to its organic form, but not to Corbusier. Thus, this technique enables us to understand the focus of the machine’s eye for the classification of objects.


Top-1 and top-5 accuracy analysis

Grad-CAM analysis

An interior photo of Alvar Aalto’s Vyborg Library was fed into the trained model. The prediction of the top four categories is as follows: Aalto, Gehry, Ban, and Tschumi. In this example, we can observe that the skylight window was the main reason for the model to pick up Aalto as its top choice.

An exterior photo of Alvaro Siza’s Porto School of Architecture was fed into the trained model. The prediction of the top four categories is as follows: Siza, Tschumi, Hadid, and Pei. By using Grad-CAM, we were able to observe the evidence of the machine’s eye’s focus in each image and the reason the computer vision made these decisions, with the probability for each choice.

PCA and clustering analysis

For example, the first cluster consists of Norman Forster, Richard Rogers, and Renzo Piano. They are frequently labeled in terms of “high-tech design”, which pursues the expression of technology (i.e., structure and facilities) as a design elements. The fourth cluster consists of Enric Miralles, Peter Eisenman, and Tadao Ando. Although Miralles was not originally classified as a “deconstructivist”, the characteristics of his architecture can be described as fragmented, inclined roofs and walls, and are seemingly under construction, which is similar to the architectural characteristics of Eisenman, who is classified as a “deconstructivist.”


Yoshimura, Y., Cai B., Wang Z., Ratti C. (2019) Deep Learning Architect: Classification for Architectural Design Through the Eye of Artificial Intelligence. In: Geertman S., Zhan Q., Allan A., Pettit C. (eds) Computational Urban Planning and Management for Smart Cities. CUPUM 2019. Lecture Notes in Geoinformation and Cartography. Springer, Cham

▲ Back to Top