Neural style transfer uses a pre-trained neural network to map the “style” of an image onto the “content” of another image. When executed between two plans, this method literally reimagines a “content” plan in the geometrical style of the “style” plan. It will be used to cross-breed architectural visuals (facades, plans, renderings, sketches etc.) of various styles, e.g., reimagining the plans of van der Rohe articulated in the style of bulky Baroque plans. These investigations provoke deliberations over the role of the machine as the “doer”, the “thinker” or the “creative companion”; and the design implications of cross-breeding artefacts of different styles.
This 2-day part of the course will introduce the use of Rhino & Grasshopper scripting as an aid in design exploration by looking at design as a complex inter-relationship of parameters. Complex surface modeling in Rhino, and algorithmic thinking and parametric tower design script will be introduced in Grasshopper. Based on the parameters, different options will be explored, by automating the process of exploration, similar in principle to the rapid doodles developed during initial conceptual stage. The technique of parametric design marks a paradigm shift where relationships between elements are used to manipulate and inform the design of geometries and structures. The focus shifts from the process of modelling geometry, to the process of defining geometry.
This exercise will be an introduction to the training of a neural network with the use of a precollected objective data set. Since this course is designed to be specific to the domain of architecture and design, the data set will be housing costs as a function of input variables such as area demographics, property facilities and structural age. On training a neural network with the data set, the trained neural network will be used to predict housing cost of areas or properties whose data is not available or collected. Through the process, participants will get a hang of how to use csv files with machine learning, how to read and generate data relationships, and how to export the new predictions from the trained neural network.
In this problem, the age-old question of “How to quantify subjective opinions of design?” will be addressed. As an example, a computational methodology will be used to quantify the visual perception of originality of design using crowd-sourced evaluation data. A parametric model of Absolute Tower by MAD Architects, defined by twelve variables shall be used to generate a few design options (a simpler parametric version of the tower will be covered in the Rhino and Grasshopper part of the course). Subsequently, using appropriate data gathering forms, participants will be asked to rate the options on a scale of ‘plagiarised’ to ‘original’, in comparison to the original design of the Tower. With the information of the limited set of design options, a neural network will be trained to map the level of originality from the limited set of design options to all possible design options. Use of neural network opens the possibility of assessing design options generated by varying multiple parameters (12 variables in the example). It is a paradigm shifting leap compared to singular or double parameter variations that the human brain is capable of perceiving and calculating. The methodology can be later used by the participants to assess any subjective opinion, e.g., urban safety, design beauty etc.
Urban perception has long been studied from the perspective of active measures (such as patrolling, adequate lighting at night, and fast grievance addressal systems for urban safety). The use of machine vision allows designers to interrogate how urban fabric affects such perception, and consequently design interventions that may increase perceived safety. In this exercise, machine vision algorithms such as semantic segmentation and object detection will be used to extract and quantify urban features (e.g., road, people, trees, vehicles, sky, building, etc.) from photographs of urban scenes. This exercise will equip participants to later on (post the course) get the same photographs reviewed to get subjective perception scores such as safety and beauty. Participants can use the methodology from exercise 3 to subsequently predict urban safety or beauty of new photographs using the quantified features in conjunction with the subjective reviews. Use of machine vision and machine learning opens the possibility of statistical assessment of multivariate urban spatial information. Additionally, using crowd-sourced data marks a departure from the topdown evaluative guidelines published by experts to a more inclusive bottom-up evaluation by end users.
Participants are expected to gain the following-
| Topic | Focus |
|---|---|
| Concepts of AI & Machine Learning | Introduction |
| AI exercise 1 : Multi-style Transfer using pre-trained Neural Network | Style transfer exercise |
| AI exercise 1 continued | |
| Introduction to Rhino : complex surface modeling | Rhino & Grasshopper |
| Introduction to Grasshopper : basic concept + parametric tower | |
| AI exercise 2 : Training Neural Network to predict Housing cost | Machine learning for objective prediction |
| AI exercise 3 : Quantification of subjective evaluation * | Predicting subjective evaluation using Machine Learning |
| AI exercise 3 : Crowd-sourced evaluation data collection | |
| AI exercise 3 : Training Neural Network to predict design Originality | |
| AI exercise 4 : Urban image segmentation into constituent elements | AI in Urban Design |
*Evaluation can be based on any attribute that cannot be predicted with formulae, e.g., design originality, perceived urban safety, beauty etc.
Tools:
Rhino
Grasshopper
Google Colab
Adobe Photoshop
No prior knowledge of the tools is needed
