Image placeholder

Image placeholder

“The successful inventor asks where we can get from here rather than how we can get there.”

Kenneth O. Stanley & Joel Lehman, 2015

Welcome to the interactive presentation. Below you will find a complete overview of the thesis. You can click on topics of interest to move directly. To follow in sequence, simply scroll with the content. Enter through to the next room with the blue button when available.

Architectural Learning

Towards a digital activation of our past


The result of this thesis has been made possible through the weaving of many threads of thought, originating far beyond my own creation. The project is also the end of a 5-year chapter at university. As such, I would like to thank my family; my parents, and sisters, who along with good friends, anchors the context of such long endeavours of curious exploration, truly meaningful.

I would like to express my sincere gratitude to my supervisor Jørgen Hallås Skatland, for his deeply helpful contributions, and for whose previous research, as well as insight and skilful guidance, this learning process has benefited greatly. His genuine curiosity, engagement, and breadth of knowledge have been a source of inspiration.

I would also like to thank Øyvind Bodsberg from OBOS Oslo. The opportunity afforded at Ulvenveien in March became a very important stepping stone for the rest of the thesis. OBOS's collaboration and the extraction of architectural layouts from their archive gave important depth to the research and provided valuable insight.

For the helpful assistance with the interactive demonstration, I wish to extend my thanks to Kent Are Torvik. I would like to thank Øystein Dunker, for his thorough and helpful readings and feedback. I am also thankful to Jane Pernille Landa Hansen for her feedback and thoughts on the work in May. I would like to extend thanks to Pasi Aalto, who contributed with important early insights and a very helpful set of reading material. Sondre K. Knutsen and Anders Torp’s feedback and support have also been greatly appreciated throughout the project. I am also grateful for the discussions with Øyvind Eikestøl Legreid, Bjørn Mikal Svendsbøe Høyland, and Peter Wilhelm Valerius Aasgaard. Stanislas Chaillou’s master thesis from Harvard in 2019 was an inspirational pivot point to my decision of exploring AI for the master's thesis. Additionally, my work would not have been possible without the research by Phillip Isola, and the efforts of Christopher Hesse, of which he generously shared with the public.

Lastly, I would like to extend thanks to NTNU and those at the university that have, throughout these five years, allowed us to build understanding and knowledge from a platform rooted in deep traditions. Being introduced to the combination of study at Delft University of Technology in the Netherlands, and the years at NTNU in Trondheim, including the teachers and students who have been a part of this experience, have left me with a deep sense of gratitude.

Abbreviations and Definitions

• AI – Artificial Intelligence
• AEC – Architecture, Engineering, and Construction
• ANN – Artificial neural nets
• BIM – Building Information Model
• Building Code – The underlying model is to be understood as the combination of law and enforced regulation. Currently this includes The Planning and Building Act of 2008, Technical Regulations (TEK10 and TEK17), Building Application Regulations (SAK10), and Product Documentation Regulation (DOK).
• CGAN/cGAN – Conditional Generative Adversarial Network
• GAN – Generative Adversarial Network
• IAI – International Alliance for Interoperability
• IFC – Industry Foundation Classes
• PBL – The Planning and Building Act (Norwegian: Plan- og bygningsloven)
• TEK – Building and construction regulations focusing on technical requirements.

Chapter 1: Initiation

Architecture has been described as a reflection of our values and our technology. A cultural expression of social cohesion, power, and the collective psyche at a given time. In a continual cycle we shape our environment, and our environment shapes us.

Could it be like something, to be architecture? How should our collective architecture learn or think? How do we ensure learning from past architectural experience? This thesis is an exploration into new territory, with the hopes of stimulating conversation and opening new pathways to secure and safeguard architectural quality in a world under rapid change.

The work will seek to engage and explore novelty at the intersection between architecture and artificial intelligence. While doing this, it will also challenge large systemic structures governing and shaping our built environment. The work is done with the ambition that we may one day lift the processes that constrain, shape, and evolve our everyday architecture, to a level also embodied by our current knowledge on learning.

1.1 Background and Motivation

The following work is grounded in questions concerning the role and responsibility of the 21st century architect. It has also been motivated by a deeply felt curiosity to continue questioning the most fundamental definitions of architecture itself. We have, for the first time in recorded history, become a global civilization. This entails problems that can only be solved by global consensus and by globally implemented solutions. We know that civilizations rise and fall, but now that we are one, it is more important than ever to face the reality of our challenges quickly. It beckons society, and should energize architects, to search and explore new paths and new opportunities, to solve the difficult problems ahead. For many of us, the familiar phrase "think globally, act locally", as attributed to the town planner Patrick Geddes, captures a practical and useful sentiment.

A well-known fact amongst architects, planners, and builders, is also the immense effect of the building code on much of architecture and on the built environment in general. Presumably by design, it has not been addressed throughout our education at university. I did, however, eventually arrive at the late realisation, that the building codes deeply shapes our work and our lives. This became an important complementary factor to my motivation for working towards meaningful and effectful impact.

Recently I am reminded of the 1977 video “Powers of Ten”, by the Office of Charles and Ray Eames.1 The video zooms out from a picnic scene in Chicago, and in doing so, moving our perspective. In the very end we have moved 10^24 meters outside of the original scene, far beyond our Milky way galaxy. The movie then zooms back in, to the picnic scene. We now move down to 10^-16 meters, at the level of protons and neutrons within an atom of a carbon molecule. Such changes in perspective, in combination with the observable changes over time, can give rise to powerful and profound insights and ideas.

Such new ideas and insights may, in the future, be derived from the observation of our building code defined as a collective learning process. The building code, understood as a learning process was first suggested, and formulated by my supervisor Jørgen Hallås Skatland. The continued exploration of this observation has served as an important backdrop and a portal to what later became this thesis.

Regarding the topic of the architectural profession, it appears that the practical modus operandi and role of architects made a slow and subtle, but noticeable change in the 20th century. From a combined designer, planner, engineer, artist, and project leader; to something similar yet subtly different. If truly existent, the evolution is perhaps just a change in “shape”, leaving the essential form untouched. However, questions concerning our responsibility and place within the larger context of society is deeply interesting, as we have shifted into a new type of expert-society. The role of the unambiguous specialised generalist may be more important than ever.

I have previously suggested the change, if there truly is one, to be that of an understandable counter-reaction to the profound post-enlightenment consequence of an increased focus on the value of all things measurable. Seemingly, a result from our viewing of the world in mechanistic terms. In turn, a stance marked with an arguably unbalanced grip on the practicalities needed in a chief builder, have increasingly emerged. I speculate whether it may have led us somewhat astray, specifically with respects to the balanced act of taking full responsibility and advantage of the technological innovations outside our immediate discipline.

I am also curious as to the arguments against a responsibility for continual and practical exploration and appropriate updating of our toolkit. After all, the pencil is as much a technology as the 3d printer. This is not to say that we should accept and use technology simply because of its existence. Importantly, it seems that the architect must continually balance his or her stance within the measurable and immeasurable arenas of exploration. For the future, I question whether we need to stay more firmly within both the scientific as well as the artistic scene. Expanding and slightly increasing our work into the scientific, technological, and quantifiable arena may afford us greater, and perhaps needed, influence, to partake in the decisions concerning future living environments and future cities.

The consequence of not being more “forward-leaning” towards such arenas, could result in the inability to adequately understand, and resultingly communicate, new and needed architectural solutions to the broad range of social decisionmakers. A consequence of not firmly placing both our feet within the immeasurable and the measurable, could be the loss of footing within the important societal discussions of the 21st century. This could include important opportunities to direct developments where architecture can greatly increase human wellbeing. In 2018 the UN predicted that 68% of the world's population would live within urban areas in 2050.2 It is therefore an unmistakably discouraging scenario, should architects lose grip on the responsibility for human needs within the built environments and future cities.


1. Office of Charles and Ray Eames (1977), “Powers of Ten”, Accessed April 2020, Available at:

2. United Nations (2018), “2018 Revision of World Urbanization Prospects”

1.2 Focus and the Landscape of Investigation

The thesis explores the possibilities at the intersection of architecture and artificial intelligence (AI). The latter technology is at the heart of much of our daily life, often without ourselves realising it. Additionally, AI has made significant impacts to a large set of industries. However, the architecture, engineering, and construction industry (AEC) is still lagging.3 An initial area of focus before beginning the project was therefore concerned with how future architects could utilise artificial intelligence in their creative processes.

As a second step, to constructively narrow the investigation and research, the Norwegian building code has served as a focus point. The building code states the constraints, goals and boundaries of our everyday architecture and built environment. It therefore follows that working for positive changes to such an immense effector of change, could resultingly lead to proportionally positive outcomes.

The choice of focus is tied to the already mentioned implications of a research paper of interest, written by Jørgen Hallås Skatland. The paper states how the building code affords us a dynamic type of societal blueprint.4 A new observation and consequently a new understanding of our building code seemingly emerges. The building code seems very much like a collective and social learning process. As such, the building code reflects our current technology and our societal values, while importantly also forcefully guiding the construction, renewal, and maintenance of our societal structures, and consequently our physical reality.


3. McKinsey & Company (2018), “Artificial Intelligence: Construction technology’s next frontier”

4. Jørgen Hallås Skatland et al. (2018), “Society’s Blueprints - A Study of the Norwegian Building Code’s Modal Descriptions of a Building”, Nordic Journal of Architectural Research

1.3 A Few Notes on the Approach

The approach and process have been akin to a journey into unknown territory. Overall, the work should be understood as exploratory and experimental. Throughout, I have sought to allow the research findings, as they have appeared, to guide the project flow. Each finding has therefore, one after another, provided stepping stones for further discoveries and new directions of investigation.

The choice of an experimental and open approach has been particularly important as the thesis, at times, explores areas, that to the authors knowledge, have yet to be thoroughly researched. The application of constraints, rather than primary goals, has been motivated by the research and writing of Kenneth O. Stanley and Joel Lehman.5

The initial approach was therefore also marked by a certain willingness for risk. This mainly concerned the assumed, but still unknown, challenges with regards to utilising machine learning technology on the architectural topic. It was clear from the very beginning that a substantial amount of new understanding, in a different discipline, was needed. Computer programming and machine learning-tools – still being a far cry from the traditional architecture student’s toolkit.


5. Kenneth O. Stanley, Joel Lehman (2015), “Why Greatness Cannot Be planned: The Myth of the Objective”, ISBN-10: 9783319155234

1.4 Resources, Technology and Limitations

The project has been subject to the following limitations:

Time: Project duration was set from January to May, spanning a total of 18 weeks.

Resource limitations: Some resources have been scarcer than first planned. Specifically, due to the unexpected shut-down of campus in March, on the grounds of the covid-19 pandemic. This included access to a suitable and powerful computer for efficient machine learning training. Additionally, plans were made to 3D-print a range of outcomes from the project. However, access to 3D printers were also lost in due to the shut-down.

Notes on technology and tools: The project uses a specific machine learning model based on a conditional generative adversarial network.6 The specific model was first presented in a research paper from the Berkeley AI Research lab.7 The thesis does not include deep theoretical knowledge on the machine learning tools applied. Such information is openly available elsewhere, and outside the scope of this project.


6. Christopher Hesse, Tensorflow adaptation of Phillip Isola et al [2] Image-to-Image cGAN, Accessed: April 2020, Available at:

7. Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (2018), “Image-to-Image Translation with Conditional Adversarial Networks”, Berkeley AI Research Laboratory.

1.5 Exploration Overview: Questions of Interest

Image placeholder

Chapter 2: Architecture, AI, and the Building Code

“Architecture is a reflection of our values and our technology.”

Fredrik Lund, 2019

As we have established, the focus of this thesis is the intersection between architecture and artificial intelligence. To narrow the work and effort, within a productive and meaningful topic, the building code has been chosen as our exploratory landscape. In this chapter, we will spend some time building a platform to better understanding the thesis’s focus. We will explore the outcome of the initial process and research. All of which, led to our problem formulation and research questions, as seen in the exploration overview from chapter one.

Technological innovation will surely always be of interest to architects and to the field of architecture, as it combines with the forefront of structural and material engineering to broaden the possibility of imaginative, artistic, and cultural expression. As such, the subject of technology, including that of artificial intelligence, should certainly be central to the explorations of architecture and its adherents.

Of great importance for everyday architecture, we find the idea of law and the building code. “Building codes are descriptive texts that reflect societal consensus and convey societal agency on the built environment. These texts represent an available, empirical source on a societal component of the built environment, containing expressions of enforced social necessities embedded into buildings and spaces.”1 It may therefore come as no surprise, that the consequences of these codes are, as I will elaborate on soon, immense. This pertains to both the positive consequences of well implemented updates, and likewise, the often unforeseeable and tremendously costly consequences of poor updates. Having chosen our initial search space, let us begin by looking at the empirical observation that sparked the process.


1. Jørgen Hallås Skatland et al. (2018), “Society’s Blueprints - A Study of the Norwegian Building Code’s Modal Descriptions of a Building”, Nordic Journal of Architectural Research

2.1 The Building Code: A Learning Process

Importantly, a large part of the work in this thesis is built upon the empirical observation that our building code is a learning process. What do I mean by this?

The building code, when viewed at a moment in time is static and unchanging. However, such a reading is arguably misleading with respect to its meaning. Its context is not only its physical location and area of effect, but also its place and being over time. To render meaning, we must therefore place it within its proper context.2 When seen also through time, which is the more appropriate contextual setting that it operates within, the building code is clearly dynamic, with regular updates to its own model.

For every update we will either add new content, or we will remove old content that we deem unfit. A new update will thereafter lead to new outcomes, that leads to more understanding, and we thereafter update again, in the hopes of implementing the lessons. These updates, or control adjustments, are done in accordance with our perceived effects of the current, inadequate code. Therefore, what we are really witnessing, within a different contextual perspective, is a societal and collective learning process. Within this process lies the model which lawfully expresses, and therefore forcefully guides, the requirements for structural organisation of large parts of our built environment. The structure of this learning process seemingly takes the form of a complex cybernetic loop. Below we find a simplified illustration.

Image placeholder

The key take-away from this observation is that we can now compare this learning process to other learning processes. Abstracting away from the particulars of the building code, we will quickly learn that if this empirical observation is correct; we have a genuinely sub-optimal structure for allowing the building code to learn efficiently.


2. Gregory Bateson (1979), “Mind and Nature: A Necessary Unity”, Hampton Press, Inc. 2001, p. 14

2.2 Current Vulnerability and Significance

Having made the claim that our current building code may lack an efficient learning structure, let us look at why this is important to address.

I believe that today’s structure is deeply vulnerable because we base the building code's structure, including its very important editing and updating, on few and limited means. These specifically involves the overwhelming reliance on decisions makers who will always be subject to self-serving preconceptions, limited experience, bias, and incomplete presumptions. This platform for decision-making looms behind us as we make most of our decisions in today’s modern society. However, striving to improve this platform should be a top priority. Particularly so, because the consequences of these vulnerabilities are reinforced by the immense second order effects of unsuccessful measures and updates.

As an architect I would assert with confidence that the built environment directly ties to our wellbeing. Therefore, changes to these man-made environments also directly result in profound consequences to both the individual citizens and to the national state. To illustrate this, let us briefly focus on the building code’s intervening effect on the national level.

The built environment represents both direct economic values and indirect and real social and environmental values. As such, we may glimpse a small part the major effect of poor or unbeknownst socially destructive updating of the building code – as well as the excellent opportunities for positive change – by looking to the economical values associated with our buildings.

According to the Ministry of Local Government and Regional Development, 66% of Norway’s “realverdi” (of which translates to English partly through the expression “real value”) is made up of our built structures. The ministry have consequently put the following official statement forward, on the topic of our building code: “It is therefore of fundamental significance to have good and effective building- and case proceedings, and a building code that secures lasting and good quality for housing and buildings.”.3

According to the last calculation of the Norwegian national wealth, completed in 2016 by the Ministry of Finance, the real-value capital for each citizen is set at 1.9 million NOK.4 There has been a strong growth in housing construction since 2016, but for now, let us use these numbers and multiply by the current population. Resultingly, a total real-value capital of 10,178,300,000,000 NOK emerges. It follows that 6,717,678,000,000 NOK, the 66% of this 2020 real-value, is consequently managed and secured in direct or indirect ways, by the building code.

We can also move one step further and address the social and environmental values. These may also be directly associated and expressed through some selected few economical terms. When we look to the national wealth studies, by the Ministry of Finance, “human-capital” is the largest post. It totals 75% of the Norwegian national wealth.4 When we apply the 2016 numbers to today’s population, this accounts for an economical value of 56,248,500,000,000 NOK. We should also be mindful of the immeasurable social values associated, of which we will have a hard time converting to economical statements, or measurable and precise numbers; even though they, in the strongest sense of the word, carry value.

Having established an account of the immensity of the value placed within the human-capital, we can further focus on the degree to which the humans within the human-capital are affected by their environment daily. Particularly, with respect to wellbeing and therefore productivity. According to Statistics Norway (SSB), and their time use-survey from 2012, Norwegians spend, on average 21.5 hours a day inside.5 This accounts for around 89.6% of the time spent inside the built environment, again, every day. It is therefore no high claim that the structures themselves may also be held largely accountable for the effects on human-capital. Not only do the built environment affects us while we are inside buildings; we also know they meaningfully co-create the outside urban spaces that we move in when we are outside in cities and urban areas.

The consequences of good natural lighting, adequate air quality and temperature, adapted noise regulation, functional layouts for work, social interaction, and calming atmospheres, are well known. The holistic design of all such aspects and many more – in short aesthetic, constructive, and high-quality incoming stimuli from our environment – profoundly matters. I am in no way making a new argument here. However, as exemplified through this economical perspective, the building code deeply intervene in these tremendous economical assets of national wealth. I therefore claim that the significance of the building code’s impact is hardly understated. As we multiply the vulnerabilities with the effects of its likely and possibly necessary updating mistakes, the importance of continually researching and improving its learning capability therefore becomes of increasingly great significance.


3. Ministry of Local Government and Modernisation, Accessed: February 2020, Available at:, English translation by author: Original: «Det er derfor av grunnleggende betydning å ha gode og effektive bygge og saksbehandlingsprosesser og et regelverk som sikrer en varig og god kvalitet i boliger og bygg.».

4. Ministry of Finance, Available at:

5. Statistics Norway (SSB) (2012), Accessed: February 2020, Available at:

2.3 Learning Efficiency: Looking to Neuroscience

So far, we have seen that our building code may be understood as a learning process. I have also claimed that the current state of its learning structure is sub-optimal. This is, for example, illustrated in the way the building code misses necessary self-rebalancing, as political policy and markets slowly turn housing into instruments of investment. It can also be seen when the very core structure indirectly starts incentivising intelligent outside agents to optimize or “game” the current content. Both such actions directly contradict its main mission and function (Part I: General provisions / “Formål” §1).6

I believe many of these effects are consequences of its inherent abstract architectural composition, which afford only sub-optimal assistance with regards to the available information processing and updating premises. Resultingly, the improvement of laws and overall learning outcome over time suffers. One way to advance our understanding of learning, is to look to neuroscience. A field that has seen a large increase in research7, and that could seemingly hold great value and interest to the field of architecture and architects.8

From Karl Friston’s work on the human brain, and its active inference and learning, we may discover more optimal characteristics for learning. One such characteristic is the ability to always receive and process input with high efficiency. By receiving, analysing, and structuring as much as possible of the signals from the outer environment – and keeping closely in touch with the results of one’s own decisions – more input may be converted to useful information on which to further update action policies.9 With an increased ability to receive and process signals of input, these units of potential value may be compressed to information, and further compressed to knowledge, resulting in a greater platform for decision making. According to work by Jürgen Schmidhuber, we may ourselves effectively be driven by such internal rewards related to our ability to progressively compress, and in effect better understand, particularly in relation to energy use, the signals and information we absorb and utilise.10

It will perhaps come as no surprise, that our embodied brain’s ability to continually receive and process signals from its environment is a central key for learning, and subsequently, a key for achieving success in its vital ability to model and understand the outer world. This modelling happens through interacting with the world, making the world more like its own model, or through being affected by the world and in turn change itself.9 An important point to our new understanding of the brain and its embodied cognition, is that we are slowly thinning the boundary between the outside world and ourselves. The clear split is becoming less so.

Another central mechanism for effective learning, is a well-developed system that can continually receive and apply feedback concerning deviations from expected occurrences. “What did not happen the way I predicted?”. These deviations from our expectations, or better understood as surprises, is according to Friston, a central focus point for learning. The continual minimising of surprise, and therefore optimising of free energy, through more accurate updating of one’s own model of the outside world, seem then our essential hallmark for effective learning.

Since 18.06.1965 the main building code (PBL) has been updated 48 times.11 This slow updating process12 (a new update, on average every 420th day) is completely understandable in a time where the structuring capabilities, technological systematisation, and predicative data analysis could not afford the opportunity to better understand the effects of the updating itself. I also emphasize with the sense of security and stability provided by a slow pace, given the available context of the past. This thesis is however an exploration into possible futures, and even today it should be no far stretch of the imagination to see a road towards immediate, and large-scale, feedback and analysis.

A sudden and historically unprecedented challenge is suddenly having access to more information than we know how to properly structure. Returning to the specific topic of the built environment and construction; we are already moving towards building digital twins /BIM-models of both new and old projects. The latter being reasonable to assume because of our digital working methods, as well as our ongoing need for continued upgrading and maintenance of the old building stock. These digital twin models could take part as signals for a learning mechanism that continually receives feedback on how the law shapes new consequences and new surprises.

The difference in outcomes, from a building code structured on sound learning principles, as opposed to the current situation could be illustrated by the following analogy concerning instruments. Given that you were to learn to play the piano, how fast would you learn to play, if the sound from your keys made their noise heard 420 days after you pressed them; compared to hearing them immediately?


6. Ministry of Local Government and Modernisation, Plan and Building Act 2008, Accessed: April 2020, Available at:

7. Yeung AWK, Goto TK, Leung WK (2017), "The Changing Landscape of Neuroscience Research, 2006–2015: A Bibliometric Study", Frontiers in Neuroscience, 11:120. doi: 10.3389/fnins.2017.00120

8. For those interested in pursuing the subject I would recommend the "Mind in Architecture - Neuroscience, Embodiment, and the Future of Design" (2015), by Sarah Robinson and Juhani Pallasmaa, MIT Press, ISBN: 9780262533607

9. Karl Friston et al. (2016), “Active Inference and Learning”, Neuroscience & Biobehavioral Reviews, Volume 68, Pages 862-879

10. Jürgen Schmidhuber (2009), “Driven by Compression Progress: A Simple Principle Explains Essential Aspects of Subjective Beauty, Novelty, Surprise, Interestingness, Attention, Curiosity, Creativity, Art, Science, Music, Jokes”, Journal of SICE 48(1), 21-32, 2009, arXiv:0812.4360

11. Ministry of Local Government and Modernisation, Complete Overview Plan and Building Act 1965-2017, Accessed: January 2020, Available at:

12. Compared to the human brain/ the most advanced and intelligent naturalistic learning mechanism we are currently aware of.

2.4 A Digital Age on the Rise

Having briefly explored the idea of looking to other efficient learning systems, let us reflect on why these ideas mentioned so far, could be powerful or even relevant. An answer to this may be found by briefly sketching parts of our current place on the timeline, and the resulting environment we find ourselves in.

At this moment, close to 4 billion people on earth have nearly the entire knowledge ever accumulated available at their fingertips. Since around 2008, and the advent of the first successful “smart phone” (essentially a computer with dimensions fitted for always having it on your body), we have, without much fuss, merged quite successfully with our odd digital flat pocket-size devices. Our relentless and seemingly unstoppable use of these digital interconnected devices also leaves a staggering amount of data and digital signals available, for individuals who wish to learn more about human behaviour and the world at large. How much of our life are we converting to bits? “As of May 2019, more than 500 hours of video were uploaded to YouTube every minute. This equates to approximately 30,000 hours of newly uploaded content per hour.”13 If we look to a more general metric, we find that the data production in 2018 was 2.5 quintillion bytes a day.14 This pace is continually increasing. At the same time, we are faced with just the right technological developments to start structuring and understanding volumes of this complexity. With the advent of immense data and large-scale statistical tools, we are starting to construct order out of this seemingly unorganized bewilderment of connectedness.

We may also briefly digress, to point out that we are, today, in effect, moving larger and larger parts of our consciousness into the construct of digital space. Should it concern architecture? To what extent is it our responsibility to explore this merging of the worlds of atoms and bits? We will leave these questions for now. What is certain, however, is that we are building a new type of infrastructure on top of the old one. An infrastructure reflecting the rise of a digital age. This platform will increasingly allow us to structure incoming flows of signals from our environment, and recursively implement improved learning in the new age permeated by the carbon-silicon environment and the human-computer interaction. It is reasonable to assume, that efficient learning meaningfully connects to adaptability. As natural selection would have it, those given the best fit to understand and continually adapt to their changing environment will leap ahead. Let us therefore look further into the new opportunities, provided by data and these new infrastructures. We will begin by building a brief platform to better understand the statistical tools, of which can harness the fuel of a digital age.


13. J. Clement (2019), Statista Article, Accessed: April 2020, Available at:

14. Bernard Marr (2018), Forbes Article, Accessed: April 2020, Available at:

2.5 Artificial Intelligence: A Very Brief Introduction

The impact from artificial intelligence (AI) on the future is expected to be profound.15 However, it would probably be fair to say that AI, for most us, is somewhat clouded behind the complexities of its operational mechanics, its popular hype, and possibly our lack of understanding with regards to how it ties in with our daily life. Still, the subject has slowly been introduced to a wide variety of visions for our future. A future where this technology places a central part. Assuming there is such a thing as artificial learning and intelligence, let us briefly explore the topic.

There is no lack of science fiction movies (a medium and genre that may be a major shaper of our current history) to vividly express the dystopian and utopian promises of our topic of choice. A large set of prominent authors and public figures have also passionately expressed grave concerns, or heartfelt optimism, at our collective AI-enriched future. Indeed, we need go no further than the prophesies of Elon Musk16 or the underlying message in the books of Ray Kurzweil17 or Nick Bostrom18, to start wondering how this technology might be impacting our future. But what is it really?

AI, or sometimes called machine intelligence, is the term we use for intelligence demonstrated by machines. In some ways, it is that simple. However, this of course instantly begs for an intelligence-definition (of which we would find no lack of suggestions), but let us stay on AI generally, for now. From another perspective, we may say that AI and its existence and development is a pure reflection of the human curiosity, and its deep hunger for self-understanding. This may be argued from the fact that much of AI is inspired by our continued understanding of the human brain, particularly its neuronal structure. However, as many neuroscientists with some knowledge of artificial neural networks might say, the comparison is a far from being very good.

The academic discipline was founded in the 1950’s and it is today divided into a large set of sub-fields.19 One of which, this thesis will pursue further, called machine learning. At the core of this technology lies the somewhat timeless tool and language of mathematics, in combination with the more recent field of computer science. One may also say AI is intelligent application of mathematics, and more specifically statistics, in combination with computer algorithms. To this point, we could describe AI as computational architecture, of ordered and logical structures, that facilitate imitations of our very fundamental human cognitive functions. This breaks down to the ability to receive, analyse, and act on a signal of communication. At heart of the analysis is the ability, on its own terms, to tell different things apart from each other. When this is done in a system of relations to a set of cleverly structured recipes of action – what we may call a code, or an algorithm – the results can be proportionally amazing to the human creativity.

What makes AI so powerful, is exactly this ability to replicate fundamental cognitive abilities, and partially create its own understanding of the input. Essentially, this achievement seems closely linked, and is often therefore paralleled, to the story and evolution, that eventually became our own human unfolding. Today AI exhibits greater than human intelligence in a large range of activities. In 1997 the world’s then number one chess player Garry Kasparov, lost to IBM’s Deep Blue chess playing machine.20 A milestone of achievement later to be surpassed in 2016 by the Alpha Go-machine’s victory over Ke Jie (the world’s then top Go player).21 Today different technology, utilising artificial intelligence, is branching out. It is reaching a greater and more creative audience, and new development and achievements have clearly been made that should interest designers.22

However, as with all technology, some of the innovations are also clearly holding dangerous potential right out of the box. As an example, everyone may now utilise automatic re-animation and motion modelling, from a simple photograph of someone they wish to impersonate.23 This lends almost endless opportunities to create realistic fake pictures and animations. AI technology has also been developed to understand soundwaves, in order to recreate voices based on limited input data. As always, technology offering opportunities for both good and bad is nothing new. We will necessarily return to this later in the thesis.

The field of AI is currently under tremendous expansion and is seeing developments and innovation at an impressive pace. In 2018 the AI software market revenues worldwide were over 10.1 billion US dollars and the market growth in 2019 was 154%.24 Looking to the future it is also reasonable to assume that a sizable proportion of this trend is likely to stay. For 2030, AI’s contribution to China’s GDP is expected at 26.1%.25 With thoughts to our immediate future, we may also look to the Gartner hype-cycle. Here, many ai-visions are seemingly still far from our doorstep. However, many more are soon ringing the doorbell, and a sizeable amount of AI technology is already “in-house”.26

As previously mentioned, an AI sub-field of particular interest is machine learning. “Machine Learning is the study of computer algorithms that improve automatically through experience.”27 We may crudely divide these types of algorithms into three categories. Those where the algorithm learns without supervision, i.e. we give no clue as to what we are asking it for. The second, to give precise directions, i.e. completely supervise. And lastly, one can allow the algorithm to learn under the structure of rewards and punishments. The latter is described as reinforcement learning. In chapter 4 we will go deeper into the structure and mechanism of a particular type of machine learning model that we will be using for an experiment into architectural learning. This model uses an algorithm consisting of two adversarial neural networks, in order to continually improve the generation of novel pictures, corresponding to an input picture database.


15. OECD (2019), “Artificial Intelligence in Society”, p.3

16. On the comment of “humans being the carbon-based biological start-up disk for artificial intelligence”, and “Summoning the demon” – The Guardian Article, 27. Oct 2014, Accessed: April 2020, Available at:

17. Ray Kurzweil (2006), “The Singularity is Near: When Humans Transcend Biology”, ISBN-10: 0143037889

18. Nick Bostrom (2014), “Superintelligence: Paths, Dangers, Strategies”, ISBN-10: 0198739834

19. Pamela McCorduck (1979), “Machines Who Think”, ISBN 1-56881-205-12004, p.424

20. Hsu, Feng-Hsiung (2004), “Behind Deep Blue: Building the Computer that Defeated the World Chess Champion”. Princeton University Press. ISBN 9780691118185.

21. Cade Metz (2016), Wired Article, Accessed: April 2020, Available at:

22. Examples: Taesung Parket et al. (2019), “Semantic Image Synthesis with Spatially-Adaptive Normalization”, GauGAN Interactive Demo: Artistic web-cam rendering, “Learning to See: Gloomy Sunday” by Memo Akten, Accessed: April 2020, Available at:

23. Aliaksandr Siarohin et al. (2019), “First Order Motion Model for Image Animation”, NeurIPS. Note: For an informative and short introduction, presented by “2 Minutes Papers”, Accessed: March 2020, view video available at:

24. Shanhong Liu (2020) Statista Article: “Artificial Intelligence – Worldwide, 2020”, Accessed: April 2020, Available at:

25. Ibid, Accessed: April 2020, Available at:

26. Laurence Goasduff (2019), Gartner Article – “Top Trends on the Gartner Hype Cycle for Artificial Intelligence”, Accessed: April 2020, Available at:

27. Tom Mitchell, McGraw Hill (1997), “Machine Learning”, Accessed: March 2020, Available at:

2.6 AI & Architecture: A Short Timeline

The concept of applying computational science to the field of architecture is by no stretch of the imagination new. As we direct our attention towards the field of architecture, we will be met with an early onset of ideas on machines that learn about architecture themselves. Let us therefore briefly explore a short timeline of this early development.

We may begin by returning to the late 1960s, and to Nicholas Negroponte, a former professor of architecture at Massachusetts Institute of Technology. Negroponte led “The Architecture Machine Group”, releasing a book on the project in 1970.28 He also mentions the idea of machines that learn about learning about architecture in his 1969 article "Toward a Theory of Architecture Machines".29 Negroponte envisioned a future where the architect and the machine would work hand in hand, in dialogue, through the design process. “His approach was significantly influenced by recent discussion on artificial intelligence, cybernetics, conversation theory, technologies for learning, sketch recognition and representation.”30 In 1973 his lab developed the research project “Urban 5”, a development on their older “Urban 2”. This machine’s original goal was to “study the desirability and feasibility of conversing with a machine about environmental design project […] using the computer as an objective mirror of the user’s own design criteria and form decisions; reflecting formed from a larger information base than the user’s personal experience”.31

A continuation on the topic of AI and architecture also emerged around 1976, with Cedric Price and John Frazer’s intelligent building project.32,33 With the "The Generator Project", the two architects sought to create a building that would intelligently adapt, through means of computer technology. A precursor for such projects would seemingly be Christopher Alexander’s 1964 release, titled “On the Synthesis of Form”. Alexander, and his ideas on pattern languages, would play an important role both in the field of architecture, as well as sparking profound movement within the computer programming discipline.

Throughout the 1980s and 1990s, renewed interest in artificial intelligence slowly emerged. A longer period of halted interest and funding, aptly named the AI-winter, was slowly coming to an end. With the increase in computational capacity, the 21st century would open a vast landscape of new opportunities.

As for current developments, a few names emerge in the public AI and architecture scene. Worth a mention here is the Norwegian company Spacemaker. The Oslo and Boston located company is currently utilising AI with the aim of improving urban site development. Many architects and researchers are also combining AI with parametric design. Using tools such as Grasshopper with AI technology, projects have for example researched methods to improve and explore more holistic manufacturing and design processes.34 Today architects, urbanists, and planners also research machine learning opportunities to map and better understand urban spaces.35 Using computer vision, an AI may “ride” Google’s street-view, categorizing and structuring frozen moments of information in our urban spaces. All the while making inferences on how architectural style is linked to real-estate prices, or how urban scenery could be connected to perceptions of beauty or wellbeing.36,37,38


28. Nicolas Negroponte (1970), “The Architecture Machine: Toward a More Human Environment”

29. Nicholas Negroponte (1969), “Toward a Theory of Architecture Machines”, Journal of Architectural, 1969 Education, Vol. 23, No.2 March 1969, p. 9

30. Eliza Pertigkiozoglou (2017), “1973 - Nicolas Negroponte and Architecture Machine Group MIT”

31. Nicolas Negroponte (1970), “The Architecture Machine: Toward a More Human Environment”, p. 71

32. Interactive Architecture, Bartlett School of Architecture, UCL, Accessed: April 2020, Available at:

33. Eliza Pertigkiozoglou (2017), “1976 - Cedric Price & John Frazer”

34. Martin Tamke et al. (2018), “Machine learning for architectural design: Practices and infrastructure”, International Journal of Architectural Computing.

35. Stephen Law et al. (2017), “An application of convolutional neural network in street image classification: the case study of London”, Alan Turing Institute, Tongji University

36. Maximilian Bachl and Daniel C. Ferreira (2019), “City-GAN: Learning architectural styles using a custom Conditional GAN architecture”

37. Seresinhe CI, Preis T, Moat HS. (2017), “Using deep learning to quantify the beauty of outdoor places”, Royal Society Open Science. 4: 170170.

38. Thies Lindenthal, Erik B. Johnson (2019), “Machine Learning, Architectural Styles and Property Values”, University of Alabama, University of Cambridge

2.7 Summary

• From a contextually meaningful perspective, the building code is a learning process. This opens new ways of understanding and structuring it.

• The current building code’s learning structure is seemingly, and unnecessarily, vulnerable and inefficient.

• AI utilizes logic based on the human brain through essentially statistical methods. The structure allows for intermediary parameters that generate novel and seemingly non-deterministic solutions.

• Projects that connect machines and machine intelligence with architecture have been topics of exploration since at least as early as the 1960s.

• As we are entering the digital age, the opportunities for implementing artificial learning to architecture are seemingly growing greater every day.

“Information is a difference that makes a difference”

Gregory Bateson, 1972