MACHINE LEARNING
Machine Learning
when
2020 . Spring
where
Cooper Union . Manhattan . NY
Ben Aranda
what
instructor
Research
Jihoon Park, Maren Spyer, Jessie Basset, Willam Smith-Clark
-
Perception of architecture is primarily driven by vision. The visual aspects essentially determine the mood of the space. For this reason, it is questioned whether a machine can possibly perceive and judge these aspects of architecture. Will they be able to suggest to us what can be used in the real world?
mjm press + Conscious CAD
-
Gan stands for generative adversarial network. It is a framework for deep learning where generators and discriminators compete with one another to create fake data. The project aims to generate new architectural data through training of collective data.
200 articles and 350 images that describe the terminology ‘deconstructivism’ were used as collective data. As a result, the fake article, written by the program, grammatically does not make sense in some parts. The image does not necessarily correlate with deconstructivism either. However, considering how close the experiment was to being successful, the possibility of architecture being automated by the program was seen.
The ‘conscious cad’ is an application that is designed for automation of plans. It generates plans through descriptions of the space by the users. Trained with over 350 plans, captioned with descriptions for each, the application provides suitable drawings for each user. Automation of drawings can be the future of architecture.
Fake Furniture
-
This project aims to build a machine learning application that could provide users with a manual of fabricating simple furniture. Using only 2X4s, the application was trained with three different data sets. Each data is a description of the same information in a different form.
The first set is consistent with texts which describe the amount and length of every element needed to be produced. Second set is compiled with an orthographic projection of three sides for each piece of furniture. Lastly, the set is bundled with axon images. Expected outcome of the prototype was to get three types description when fed with a simple sketch of a furniture.
The video below is the inner scope of the latent space where we could get a glimpse of how the machine is learning through the data provided. It is a walk through the latent space. Clearly not a form of furniture but somewhat a halfway form of getting there.
