Designing the Latent Space
First and second supervisors
The Latent Space is characterised as the hidden black box at the heart of machine learning, driving the hallucinatory images of generative art. But far from unknowable, this intelligence is a complex multi-dimensional space of semantic superimposition, compressing a dataset, and explored through vector drawings. To architects, vectors are the ubiquitous currency used to create geometry, but never before could directly calculate semantics or index features.
Architecture can utilise these latent semantics as a new method to organise and create physical space. Through the coding of networks (GAN / VAE / Transformers), custom datasets, 3D and VR environments, this research offers a radical reappraisal of the use of deep learning, not as a multi-use instrument, but to train a single spatial outcome across latent layers. Bespoke crossmodal datasets created by the architect for a project can entangle semantic, aesthetic and environmental features into one architecture of similarity and difference.
Critical research will consider the value of the latent space, situating the uncertainty, indeterminacy, and entanglement inherent within deep learning in relation to the philosophies of Clark, Boden, Parisi and Mackenzie, the ambitions of cybernetics, connectionism and information theory. Drawing upon the history of concepts of space and artificial intelligence, it recalibrates how deep learning can be used as a generative process, and whether this might anticipate a new kind of relationship between architecture and computation. Aligning spatial theories with technical studies will offer new design practice and value criteria for future AI interdisciplinary research.
Tom Holberton is a Lecturer and researcher at UCL, teaching “Unit 21” (https://unit-21.com) for ten years on the Architecture MArch (Part II) and BSc (Part I) programmes, alongside technical and design thesis modules.
A graduate of The Bartlett School of Architecture (BSc, DipArch and MArch), he won the RIBA Bronze Medal, RIBA Serjeant Award for Excellent in Drawing, RIBA Skidmore Owings Merrill Foundation Scholarship, Victor Chu Prize, Hamilton Prize and Faculty Medal.
He is a chartered architect (ARB / RIBA) and from 2006 an Associate for Rick Mather Architects / MICA for 10 years, leading a range of award-winning complex cultural and heritage projects, affordable housing, masterplanning and urban design for TfL and London local authorities. Collaborations include with SO-IL, Junya Ishigami, RCR Arquitectes and projects in London and New York.
SoHoKo was founded in 2013 as research-based studio focused on installations and exhibitions. Clients include the Victoria and Albert Museum, British Museum, Japan House London, Grosvenor, Gifu and Ibaraki local prefectures, NITTO, Asahi, and MUJI. He is an advisor and judge to the D&AD Awards.
He has developed Arts Council funded creative projects connecting Japanese craft, tea, artificial intelligence and computation.
Tom Holberton is a member of the UCL AI in Education working group, considering the impact of new AI technologies within Higher Education.
UCL Graduate Research Scholarship (GRS)
Image: Latent space sampling - Tom Holberton (2023)