Immanuel Koh is an architect and programmer, whose research works investigate creative A.I. applications in urbanism, architecture, design interaction & art. He graduated from the Architectural Association (AA) London and has taught at the AA, Royal College of Art (London), Tsinghua University (Beijing), Strelka (Moscow), Angewandte (Vienna), DIA (Bauhaus Dessau), Harvard GSD, UCL Bartlett and many others. He has exhibited internationally, such as at London’s Victoria & Albert Museum, Shanghai’s 3D Printing Museum and Taipei’s Tittot Glass Art Museum; and published widely, such as in Architectural Design (AD), Computational Design (Tongji), Design Computing & Cognition (DCC) and CAADRIA. Immanuel has also practised as an architect at Zaha Hadid Architects and as a programmer at ARUP London with Relational Urbanism. Other research grants and awards received include ARUP Research Challenge and Lars Lerup Prize. He has worked as a research scientist, in collaboration with MIT’s CSAIL, at Singapore’s SUTD Artificial Design Lab. Currently, he is a PhD candidate based at École polytechnique fédérale de Lausanne (EPFL) in Switzerland, where he teaches at the School of Computer Sciences and the Institute of Architecture.
PRESENTATION TITLE: Architectural Sampling: The Machine-Learning of Architectural Spatial Forms
The spatial and formal conception of architecture, and thus its modes of design perception and representation, directly contributes to its machine-learnability; and consequently, its capacity in leveraging today’s machine learning apparatus and deep learning models for design innovations. In order to make the ‘digital’ of architecture operative again in the context of contemporary A.I., it needs to be reframed as one that also concerns its own machinic encodings. Yet, prior to this, more fundamentally, is a conceptual leap towards an architecture that is less rigidly ontologically categorized and less precisely geometrically delineated – an overcoming of architecture’s own longstanding set of underlying dualisms, such as figure/ground, objects/fields and parts/whole. In this talk, I would like to recast architecture as a statistically fragmentary construct – a partial-object-field that can be quantized, vectorized and convolved; for the computational training, inferring and synthesizing of architectural designs. In addition to identifying historical precursors of ‘architectural sampling’, using concepts derived directly from machine vision and natural language processing, I argue for an extension of today’s parametric modelling of hand-crafted rules and associative geometries, to one of machine learning implicit spatial features and formal relations. A series of design research projects done at the EPFL (Switzerland) and the AA (London), ranging from curation of urban objects with convolutional neural networks to architectural design synthesis with predictive sequential models to infinite game levels generation with constraints propagation, will be presented to illustrate the proposed framework of ‘Architectural Sampling’.