06 A.I.R.

JAN/22 - ∞

AIR, or Architectural Image Recognition, is a result of a research project that was undertaken as part of an exploration into Artificial Intelligence and its application inside architectural practices. AIR uses deep learning to provide a ‘similarity index’ of a architectural plan against a set of ten architects. The set of architects includes Alvar Aalto, Daniel Libeskind, Jean Nouvel, Le Corbusier, Louis Kahn, Oscar Niemeyer, Rem Koolhaas, SANAA, Steven Holl and Zaha Hadid. AIR was simultaneously developed as a standalone program and as a plugin for the parametric modelling environment Grasshopper inside Rhino.

The program, in its current stage, is intended primarily towards students and researchers in architectural practice to analyze their drawings against the styles of the ‘masters’ or study the designs of these architects themselves in relation to one another. AIR can serve as a way of sparking discourse about the implications of the ‘similarity index’ while designing new projects. A more developed iteration, trained on larger data sets, can also be utilized in archiving and study of architectural drawings of various architects. It can be used in the discipline of history of architecture to find architects, or the best-guesses, of those projects whose architects have been lost to us. With a sufficient data set, and new techniques in Deep learning such as low-shot and one-shot learning, the programs capabilities can become important to the processing of architectural drawings in libraries, institutions and offices alike.

For me, it also serves as a departure point for a wider discourse on what comprises of an architects’ style and their imprint on the final building. It also highlights the limitations and mysteries of AI but also the potential that this novel technology seems to hold for a profession so ancient.

AIR is written in Python, while the deep learning model used in the program was developed through Google’s Teachable Machine which uses Keras and Tensorflow. All python libraries used in the program are available, to the best of my knowledge, as open source. The entire project was carried under the supervision and guidance of Professor Yun Kyu Yi at University of Illinois, Urbana-Champaign.