Johan Wijesinghe



→ Information
→ About
→ Email
Instagram

Domain Adaptation for
Architectural Spatial Analysis & Generation

London Years: 2019 Tags:


This dissertation investigates the potential of pre‑trained deep neural networks to interpret architectural drawings — specifically, to distinguish between plans and sections and to recognise spatial patterns. While neural networks are commonly trained on large, well‑labelled image datasets (e.g., identifying dog breeds or flower types), applying these methods to domains with sparse, irregular, or conceptually complex data — such as architectural spatial representation — presents unique challenges.

Drawing on the “projection problem” identified by cognitive scientist Jeffery Elman (UCSD), this research examines whether limited examples can adequately train networks to generalise and grasp the underlying logic of architectural drawings. It evaluates the extent to which pre‑trained networks can reduce the volume of task‑specific data needed, and explores domain adaptation — identifying where, and to what degree, a network adjusts from source to target domains.

A novel fine‑tuning strategy is proposed, using feature spaces from four auxiliary datasets — flowers, art drawings, animal sketches, and pet images. This cross‑domain approach revealed unexpected insights into how networks learn and adapt to architectural spatial logic, including which auxiliary datasets are most effective for structural and organisational understanding, and which network layers undergo the most significant adaptation.

The dissertation also introduces a new application of generative adversarial networks (GANs) to produce architectural drawings informed by pre‑trained models, demonstrating how generative processes can extend architectural dataset augmentation and exploration.

This ongoing research contributes both technical methods and theoretical insights to the intersection of deep learning and architecture, advancing how AI can be adapted to read, generate, and ultimately comprehend spatial knowledge.


ALV Kernel — An image of the first convolutional layer inside a neural network, showing how the model begins filtering and decomposing visual input for a recognition task. At this stage, the network is identifying basic contours and edge patterns rather than complete forms. Variations in tone represent the network’s response: lighter areas indicate excitatory activation, while darker areas indicate inhibition. Example inputs include a flower and a dog
Comparison of Network Layers — Visualisation of a pre‑trained neural network showing how successive layers extract progressively finer details and features. Early layers capture broad contours and shapes, while deeper layers focus on increasingly specific patterns. Similarities across layers reflect the generalised biases learned during pre‑training. Example inputs include a sunflower, a dog, a cat sketch, and an architectural CAD plan
PCA - A method of viewing how the network understands the similarity of the images after being trained. Architectural imagery on the right and architectural drawings on the left. Interestingly, sections tend to be in the middle of this diagram.



Advisors
Dr Sean Hanna - Director of Computational Design - University College London
Kahlid El-Ashry - Computational Design - Foster and Partners


Index   Next →