Associative Synthesis with Deep Neural Networks for Architectural Design: Fid-Bau Portal

Associative Synthesis with Deep Neural Networks for Architectural Design

Koh, Immanuel

Our ways of seeing can be very different from that of deep neural networks, and consequently our ways of knowing and saying what these machines are trying to show us can be serendipitously creative as a result of novel design associations. Both ways of seeing come with their own priors. For the deep neural network, it is the explicit training set used and what it sees might be visualised computationally with saliency maps. For the human, it is his/her own personal implicit perception and knowledge of design and what he/she sees might be detected empirically with eye tracking heatmaps. This paper explores such an associative interplay in co-creating through the implementation of several deep learning models that readily lend themselves to architectural appropriations. These deep models include unsupervised generative models for design synthesis, supervised classification models for design analysis and models for understanding networks’ own predictions. In architecture, the implicit and explicit association of formal ideas is one of the reasons for the in-depth study of exemplary buildings from the past and present. Therefore, the case study of buildings conducted in all architecture schools is as much an analytical exercise as it is a creative one. It trains the architecture student to see beyond that which is visible in the two-dimensional photograph or drawing of a building, and to reconstruct in his/her mind’s eye the ‘plausible’ three-dimensional spatial configuration. The paper demonstrates a series of design research projects that leverages the imagery impressions synthesised by deep generative models and the formal interpretations translated by human architects. The projects to be presented include associative synthesis between sculptures and landscape architecture, between anime film and narrative architecture, and between iconic pavilions and interpretable architecture. The proposed framework thus takes advantages of the differing capacities in learning, perceiving and synthesising as afforded by humans and machines, in order to envision new forms of human–computer co-creation. Unlike previous works, rather than having the machine to act as a fully autonomous creative agent itself a task that remains difficult and elusive even today, the paper suggests that the co-creative process is messy and requires strategic labour distribution between the human and machine, where exchanges between both entities occur frequently throughout the design process. The projects discussed also serve as samples for envisioning an appropriate co-creative framework for architectural design as well as for teaching with deep neural networks. The paper aims to answer the main research question of ‘How might deep learning models be appropriated for architectural co-creation through an associative design synthesis process?’.