Huang, Sheng-Yang; (2024) On the Architectural Rationality of Connectionist Design. Doctoral thesis (Ph.D), UCL (University College London).

Text Huang_10201420_Thesis.pdf Access restricted to UCL open access staff until 1 January 2026. Download (91MB)

Abstract

The past decade has witnessed the rise of connectionism, spurring intense exploration within the architectural research community to establish a new digital design paradigm based on the probabilistic computational mechanisms of deep neural networks. This thesis examines the conceptual shift in architectural rationality brought about by the rise of Connectionist AI and its probabilistic computational mechanisms, forming a new synthesis within the fields of architectural design, design cognition and artificial intelligence. The work rests on the premise that connectionist latent spaces provide a flexible alternative to traditional symbol-based systems, allowing compressed data to shape design schemata. This thesis considers the connectionist 'latent space' as a data compression technology where types and styles reside within. Additionally, connectionism is positioned as a pivotal shift from predetermined styles and typologies to a more dynamic, explorative approach in architectural rationality. Both these promote diversity and uniqueness in design. The thesis contains two continuous parts. The first part of the thesis provides a comprehensive historical and theoretical analysis of the evolution of connectionism within computational design methods, positioning it as a crucial shift combining architectural theory with probability theory. The second part includes case studies involving projects the author participated in, along with other selected cases, to examine the transformations within the connectionist framework across the three fields and how these shifts challenge conventional concepts of style and typology. The case studies demonstrate that the differentiable latent space enables designers to actively develop design schemata within an encoded, vector-based creative field, and to explore the space between compressed data, producing diverse designs through decompression. These demonstrate how connectionist design processes promote greater diversity and uniqueness in design, rather than leading to its homogenisation. Meanwhile, the study suggests that the errors produced in the compression and decompression processes, combined with the inherent randomness of neural network algorithms, also lead to extensive design variation. Additionally, the design approach of latent walk generates unique style hybrids, shifting style from fixed labels to temporary, unique definitions, and its composition and evolution are reconceptualised.

Download activity - last month

Export as XMLCSVJSON

Download activity - last 12 months

Export as CSVXMLJSON

Downloads by country - last 12 months

Export as JSONXMLCSV

Archive Staff Only

View Item