PlanGAN Recently, Generative Adversarial Networks have proven capable of producing high-quality extrapolations in the feature space of a given dataset. PlanGAN proposes that these developments are bringing new possibilities to design and analysis. Term: Spring 2017 Critic: Liz Diller Advisor: Jesse Reiser Duration: 12 Weeks
Academic - M.Arch Thesis
In order to make value judgments in a new drawing regime, there need to be new kinds of analysis that look closer than humanly possible. The first step in this direction was the training of an image classifier on a dataset of architectural components. A dataset of approximately 700 images was used to retrain several different top performers on the ILSVRC challenge. Densenet-121 was selected as it consistently achieved the highest validation accuracy on the 300 validation images. The best performance was just under 86%. Second retraining and verification results: ~80% accuracy.
Verification results after training with Adamax optimizer.
Verification results after training with Adam optimizer.
Visualization of the best-performing Densenetâ€™s results show the task that it is undertaking is sometimes difficult even for human experts. The images were inverted as earlier experimental results suggested that white lines on black backgrounds yielded more consistent results. The dataset was split into five classes of architectural components: columns, stairs, doors, rooms, and windows. These images show examples of the test set with their corresponding class scores from the image classifier and the correct class above the image. The images in the test set are taken from plans ranging from antiquity to the present.
The sliding window detector was deployed on a number of collected architectural drawings, both canonical and quotidian, ranging from antiquity to the present.
The five classes that the detector was trained on are color-coded in the first analysis map. A threshold of greater than 60% was used to filter false positives. This value was selected from the verification phase of the retrained detector where completely white images would return false positives as ‘rooms’ or gridded floor patterns as ‘stairs’ at or below 60%. However, the sliding window segmentation still has many pathologies. Here, highly segmented walls are returned as ‘columns’ due to the inherent inability to understand the entire context of each subset of the image.
To visualize how confident the system is, the sliding window detector creates a parallel ‘confidence map’ scored from 0 to 1.0. Here, another pathology is shown. Empty space outside the walls is returned as a ‘room’ with 100% confidence. In spite of its shortcomings, the detector does show some promising results, frequently able to detect even highly irregular stairs or find unusual interior or exterior windows.
Epoch: 5 Because of the pathologies shown in the sliding window detector, an autoencoder was ultimately used for generating the analyses. This enabled the system to discover its own features of the dataset while learning a compressed representation of the images. This compression was restricted to a 100-dimensional vector as previous experiments had shown that less than 10-dimensions was too restrictive and 100 to 1000 gave diminishing fidelity to the original images.
Epoch: 15 The autoencoder was trained on batches of 16 images over 100 epochs with stochastic gradient descent.
Epoch: 95 With this training setup, most images are able to be reproduced with good visual fidelity. Notably, many very small features are still not captured very well. Light gray â€˜fuzzâ€™ is usually an artifact of dimensions or very light floor patterns that are drawn in the base images. As the entire dataset is processed at 512x512 pixels, this loss of information likely comes from a combination of the small image size and the compressed size of the encoded vector.
Visualization of the 100-D embedded vectors of two plans. Some visual similarities are discernible, but the embeddings are not very useful without computation.
Tests of the viability of embeddings to be able to associate similar drawings showed that L2 pairwise distances of <~8.5 indicated highly similar images. In this case, the same image rotated 90 degrees scores ~5.95
Graphing the L2 pairwise distance for the embedded vector as the image is rotated through 360 degrees shows the entire range of distances.
The embedded vector proves capable of making meaningful associations beyond pixel values. Here showing that two visually and formally different plans by the same author (Hans Scharoun) are highly correlated (L2 < 8.0).
The plans dataset was then used as the source material for a generative adversarial network designed to generate images at resolutions of 1024x1024. Although these images are visually similar to architectural plans, close examination shows that they exhibit many pathologies of generative models. Many features that are key to the meaning and interpretation of plans like the symbolic encoding of doors and windows, are not present. However, the images do exhibit some amount of novelty in composition and structure. In the above image, the pattern of diagonal â€˜columnsâ€™ is a feature that is not present in the dataset and its overall form is a fairly well-resolved combination of several real examples.
This plan exhibits a similar composition to that on the facing page, but takes a very different route with its structure. Here, the drawing is defined by many sliding orthogonal walls with what appear to be underdeveloped stairs or escalators running down the central axis. Results like these, while admittedly very rough, suggest that the processing of precedents into new, yet, familiar forms is not a strictly human domain. Rather, the process is automatable and designable, and thus open for critique and improvement.
Some of the generated samples follow fairly typical constructions. The plan at left is composed primarily of ortholinear walls enclosing many rectangular spaces. The plan below, however, is comprised of features that are plausible, like a series of massive parallel walls adjacent to a field of thinner surfaces and patterns that fade in and out of focus. Drawings like these show that even this rudimentary system is capable of recreating the world similar to how it exists, as well as proposing possible, yet novel, constructions.
The facing page shows a series of drawings that are produced from a smaller dataset that focuses on a single author. Experiments like these suggested that there is a lower limit to the number of features that define a productive dataset. However, the models were consistently able to provide more extensive examinations of particular source images this way.
A general view of the tSNE plot showing a series of drawing “neighborhoods”..
The trained autoencoder is used on a set of both real and generated plans to produce embedded vectors that reduce the dimensionality of the plans from 512x512x3 to 1000 for use in a tSNE plot. This becomes a kind of formal analysis operating at many levels of abstraction that can be used to draw comparisons between disparate drawings. The image at left shows a neighborhood created from a castle, Morphosisâ€™s plans for the Emerson School in Los Angeles, and one of the generated plans. The generated images are considered to be successful if they â€œfill the holesâ€? of this plot. That is to say, its features are familiar yet novel, capable of bridging the feature space between highly diverse drawings.
This selection shows generated plans that are considered to be less successful. These have ended up in nearly homogeneous neighborhood. It is almost exclusively generated plans. This is interpreted as their having features that are too localized, making them incapable of exploring unrealized parts of the feature space.
Liberty Museum A proposal for a renovated visitor center and museum in the base of the Statue of Liberty. This scheme uses, structural, compositional, and constructive quotations as a way to propose a radically not-so-new intervention on a national icon. Term: Autumn 2016 Critic: Paul Lewis With: Kevin Pazik Role: Design, Digital Modeling, Rendering Duration: 12 Weeks
Birdseye Rendering - showing formal relationship between wrinkled facade and the statueâ€™s folds 18
Worms-eye Rendering - wrinkled facade creates reflectors for water-based radiant heating/cooling system. 19
4'-0" 3'-0" 2'-0"
Ground Level Plan
Gallery Level Plan
(5.) Anatomical Axonometric - 1. Facade assembly w/water jet-cut plates 2. Spatial trusses to support cantilever 3. Pulled-away roof with skylights/hidden truss chords 4. Underside facade with radiant pipes 5. Facade sections unrolled to be cut from sheet stock. 22
Where the cantilever crosses the existing fort wall, columns increase in density to resist the overturning moment on the rear of the truss. This creates gradual transitions from walldefined spaces to column-defined spaces.
In the cantilevered gallery, trusses are concealed on the interior to provide diverse gallery spaces. At the periphery, the trusses are revealed as they cross the angular apertures. The complexity of the facade is resolved on the interior as these varied semi-trapezoidal openings.
Section Through Atrium
Section Through Gallery Halls
Section Through Gallery and Exterior Waiting Area
Once inside the atrium, the collected sinuous surfaces separate and stretch the space of the gallery into a procession rather than a destination. This dissolves the act of waiting into a spatial experience.
The proportion of glazed to opaque facade increases from south (right) to north (left). On the roof, the angle of the wrinkled skin largely shields the skylights from harsh southern light, while allowing light in from the north.
The vertical south (right) facade has a lower proportion of glazed area. The panels that are glazed use the intense angle to completely shield the glass from solar radiation in summer months, and allow some in during the winter. On the underside, the density of the folds increases to create a radiant reflecting system that augments the covered area offered by the cantilever with partial environmental conditioning for visitors waiting in the security line.
Next Port of Call
Set in the Post Office site as one of Architectural Imagination proposals for interventions in Detroit, Next Port of Call is a speculation on new moments of urban vibrancy offered by increasing architectural and cultural contact with Detroitâ€™s waterfront.
Featured in: US Pavilion - 2016 Biennale di Architettura Venizia
With: BairBalliet Role: Drawing Duration: 3 Weeks
The Post Office
Dancing Column How does architecture change when one small piece is given some autonomy and intelligence? This project is a record of speculations around this question enacted through simulation, prototyping and design.
Term: Autumn 2015 Critic: Axel Kilian Duration: 12 Weeks
Prototype Time lapse vimeo.com/216412576
Simulation 1.0 vimeo.com/216343707
Simulation 2.0 vimeo.com/216343775
Prototype with Massing Model
Massing Model 35
Worms eye View
Worms eye Iso
Final simulations are run with robotic columns supporting a flexible roof structure. This allows the enclosed volume to expand and contract based on real-time requirements of occupancy. The convex hull of the column flock is used as one of the simulation goals. Their logic prioritizes overall stability of flexible roof and all columns. 1/3rd points of square structure. In order to preserve global stability, columns work as a group to enclose points within convex hull. Column in danger zone. When a single column is in danger of collapse from overstress or supporting too much of the roof, it calls for help from its neighbors.
The lowest-level column goal is to optimize self-stress and minimize global bending. This amounts to maximizing individual separation within roof area.
Maximum Moment Contribution
Maximum Bending Force
Minimal Moment Contribution Danger zone, Column Immobilized Overstressed
A competition entry for the Bauhaus Museum in Dessau, Vessel Collective clumps conical figures in a manner that disguises its highly ordered plan behind a shifting facade.
Bauhaus Dessau International Competition: 1st Place (tie)
With: Young & Ayata Role: Digital Modeling, Rendering Duration: 6 Weeks
Site Plan (with ground floor plan)
+4.3 UP DN
UP TO H MEC
5 12 DN
7 12 DN
Gallery Floor Plan
Section (through entry)
Deflection Analyses of Structural Frame
A plural object made from single parts. Several patterns on a single surface make multiple lines of flight while their materials surrealities eat these away.
On Exhibit: Hidde van Seggelen Gallery, London
With: Young & Ayata and Harmen Brethouwer Role: Design/Digital Modeling/Fabrication Duration: 2 Weeks
Parthole The ground is a problem. Should we touch it? Hover over it? Ignore it? Become it? Maybe itâ€™s an opportunity to create multiple performances through its multiplication. You might call this project a hermetic object, say that it has no ground, but thatâ€™s not quite right. It has at least two. Term: Autumn 2014 Critic: Michael Meredith Duration: 3 Weeks
Ground-Questioning -Reestablishing -Disbursing
What if digital fabrication could do more than approximate digital space? This project is an exploration of feedback loops between fabrication mechanisms, sensing apparatuses and the computation of materials in formation.
Exhibited in: DADA2015 (Digital Architecture Design Association)
Term: Autumn 2014 With: J. Anderson Critic: Ryan Johns Duration: 12 Weeks
1. ABB 7600 Robotic Arm 2. Custom Arduino-driven hot extrusion effector 3. Neutral Buoyancy Bath (vegetable oil) 4. Web cam 5. Laptop (receiving sensor inputs and controlling arm speed) 6. Kinect (not visible) 7. Human (with body) 8. View from web cam (compares actual deposition to target) 9. Kinect input (controls target)
The custom-fabricated end effector was designed to maximize the abilities of easily accessible arduino hobby parts. A maximum feed rate of 1.5â€?/min was calculated as the most effective for the deposition of the material. This was achieved by gearing the hobby motor down by a factor of 60:1. The reduction gearing also increased the torque to roughly 8ft/lbs. 64
Reused Hot Melt Heating Element
Laminated MDF Base Plate
60:1 Reduction Drive
Aluminum C Channel
Arduino Holder (Perforated for wifi control)
Robotic Arm Bracket
0.5m above target
0.75m above target
1.5m above target
NOthoUSE no.tho.use is made of plane-parts which assemble into a more complex whole. Each plane carries a specific formal arrangement: spaces that slip, slide, crumple and puncture each other are all equally valid.
Term: Autumn 2014 Critic: Michael Meredith Duration: 2 Weeks
Plan - Level 1
Plan - Level 2
Plan - Level 3
Plan - Level 4
By deploying a strategy of sampling from the 20th century avant-garde we propose new positions of their political forms. These phantasms are mined for new postures that are capable of populating antagonistic grounds.
Featured On: Cloudz Watching
Term: Spring 2013 With: L. Anderson, D. Grandjean, W. Hiatt, P. Miller, S. Sarver Critic: Lisa Tilder Duration: 12 Weeks
Simulationists Five of the political parties were established through analyzing the positions of canonical projects towards nature and ground. The strategy of collage was used as a way to â€œdenatureâ€? these icons to position them with and against each other.
Territories The ideas of position and posture between the parties were explored further through the creation of figured grounds which captured those forms each party deploys.
Sea of Simulation The criteria to become part of the Simulationist party were any forms that mimicked natural forms, like the biomorphic, or those that mimicked effects, like the compositional urbanism of Exodus.
Dissection by Section The section was deployed as a disciplinary strategy to dissect the forms that were created as exteriors and begin to examine the figural possibilities latent within their interiors. These insides-of-outsides became the spaces of the impossible tower.
Antagonistic Grounds The plan is redeployed as a method of evaluating territory. Colliding the insular and idealized territories with one another was mediated through The Gates which allow transference from one territory to the next.
A new space for a design school and campus plan for a historic Finnish university challenge what it means to be contextual. Through an understanding of architectural text, the park and pavilion work as foils to the adjacent schemes.
4th Place 2013 James Gui Competition
Term: Autumn 2012 Critic: Jane Murphy Duration: 12 Weeks
Proposed Condition 87
Site Model (light cannons removed)
A Box Full of Things Around It Taxonomy of Context Plan Types
Ground Floor Plan
Lateral Section 91
View from Metro Plaza
Longitudinal Section 92
Impositions A project for a performing arts center explores the definition of â€œplaceâ€? at many scales. A single object simultaneously engages its context and attempts to relate the movement of bodies through space back to the unique urban condition that it occupies.
Term: Winter 2012 Critic: Kristy Balliet Duration: 9 Weeks
Expanding the Site Regional Metro Areas:
Population: 8.4 million Avg. Distance: 85 mi
3 Major Highways 5 Major Thoroughfares Sited in prominent arts district
Outdoor rooms break open what would typically be an introverted type, the arts center, and expose it to the city at large.
Level 5 Main Gallery
Level 2.5 Performance Hall
Level 1 Entry/Lobbies
A twist on the reading room table, the highly active surface presents the work as a cohesive whole while acknowledging the particularities of each publication.
Featured on: suckerPUNCH
Team: Kristy Balliet w/ Paul Adair & Tyler Kvochick Role: Design/Fabrication/Construction Duration: 6 Weeks
Banvard Gallery Installation
After the economic crisis, Detroit is worth more empty than full; the vacant lot has come to be the defining urban type. Inverted Icon proposes an animate and articulate volume that transforms the emptiness into an active space for the city.
â€œRedesigning Detroitâ€? Competition Entry
Team: Kristy Balliet w/ Tyler Kvochick Role: Digital Modeling/Fabrication Duration: 3 Weeks
A Friendly Figure The excessive interior is proposed as the architectural mechanism that can turn a cityâ€™s blight - emptiness - from a vast anonymity to an urban volume that collects and accelerates the programs that surround it.
Percent Vacant Housing Units >15% 8% - 14%
Residential Service Office Film Incubator
Massing & Urban Room
Photos: Phil Arnold 114
Tyler Kvochick Princeton University School of Architecture M.Arch 2017