Stochastic Cities: Four Ways of Arranging 3,750 Images of Chicago

The next time you hear some­one talk­ing about algo­rithms, replace the term with ​“God” and ask your­self if the mean­ing changes. Our sup­pos­ed­ly algo­rith­mic cul­ture is not a mate­r­i­al phe­nom­e­non so much as a devo­tion­al one, a sup­pli­ca­tion made to the com­put­ers peo­ple have allowed to replace gods in their minds, even as they simul­ta­ne­ous­ly claim that sci­ence has made us imper­vi­ous to reli­gion.¹

Algo­rit­mos, the Span­ish word for algo­rithms, was the title of one of my math­e­mat­ics books in mid­dle school. I remem­ber being fas­ci­nat­ed by this term whose mean­ing or pow­er I did not ful­ly grasp: a mag­ic word or spell with which to open a uni­verse of pos­si­bil­i­ty. Twen­ty years lat­er, in an unex­pect­ed twist, the mys­ter­ies behind it became the focus of my inter­est in archi­tec­ture and design.

Accord­ing to Google Books Ngram View­er, use of the word ​“algo­rithm” peaked first in 1991 and again in 1996 (around the same year I was using the math book). Then, after a brief decline, its use began to grow, and it has been doing so at a fast pace since 2006. No day goes by when we do not hear about the impacts of algo­rithms on our lives — most evi­dent­ly, through our inter­ac­tions with online plat­forms, such as Google and Face­book. We live in a time of seem­ing­ly lim­it­less trust in these math­e­mat­i­cal process­es. We allow them to con­trol vast areas of our exis­tence, accept­ing their out­comes as if absolute truths. We treat algo­rithms as objec­tive tools, with blind faith in their math­e­mat­i­cal cer­tain­ty — a form of belief sub­con­scious­ly ingrained in human minds since the Enlight­en­ment through the ide­al of a math­e­sis uni­ver­salis. How­ev­er, the process that deter­mines so much infor­ma­tion — from rec­om­men­da­tions on Netlflix and Ama­zon, to ide­al com­muter routes, to our suit­abil­i­ty for jobs or rela­tion­ships — is any­thing but objective.

Even though the Greeks already warned us about the per­ils of assum­ing tech­no­log­i­cal knowl­edge as supe­ri­or, tech­nol­o­gy is still framed as fair and faith­ful in ser­vice to human beings, the right approach to any prob­lem we seek to resolve. How­ev­er, as Ed Finn has stat­ed, ​“aside from the most sim­plis­tic cas­es, we will nev­er know how algo­rithms know what they know.”² There­fore, trust­ing them with run­ning our world seems nei­ther ratio­nal nor wise. Yet, what Finn calls the com­pu­ta­tion­al space of imag­i­na­tion aris­es in that very uni­verse of the unknown, a place where we can chal­lenge and rein­vent our rela­tion­ship with tech­nol­o­gy. Instead of look­ing for the most opti­mized results, embrac­ing the space of the unknown allows us to see algo­rithms not as tools that auto­mate design process­es, both affirm­ing and lim­it­ing our agency, but rather as col­lab­o­ra­tive oth­ers in the con­text of cre­ative devel­op­ment. A human-machine feed­back loop that sparks imag­i­na­tion and not just the stan­dard­iza­tion most archi­tec­tur­al soft­ware allows.

Sto­chas­tic Cities explores that com­pu­ta­tion­al space of imag­i­na­tion by rear­rang­ing an aer­i­al image of Chica­go using a com­bi­na­tion of com­put­er vision and machine-learn­ing algo­rithms, not with the inten­tion of find­ing an ide­al con­fig­u­ra­tion but as an exper­i­ment to query unex­pect­ed asso­ci­a­tions. The orig­i­nal orthoim­age rep­re­sents the city from Divi­sion Street in the north to 28^th Street in the south, and from Racine Avenue in the west to the end of Navy Pier in the east. It was cut in 3,750 squares fol­low­ing the orthog­o­nal grid so char­ac­ter­is­tic of Chica­go. Each piece was then processed by a clas­si­fi­er, a neur­al net­work trained in this case on the Ima­geNet dataset, one of the first large-scale image data­bas­es avail­able online, that learned to iden­ti­fy to which cat­e­go­ry each new obser­va­tion belonged. This clas­si­fi­er ​“extract­ed” the mul­ti­ple fea­tures of each square, vari­ables rep­re­sent­ing the rela­tions between pix­els, that were inter­pret­ed and grouped by a t‑SNE (t‑Distributed Sto­chas­tic Neigh­bor Embed­ding) algo­rithm. These asso­ci­a­tions were then pro­ject­ed back onto the orig­i­nal grid.

A t‑SNE algo­rithm is a tech­nique for dimen­sion­al­i­ty reduc­tion, which is very use­ful in the visu­al­iza­tion of high-dimen­sion­al datasets. It con­sists on reduc­ing the num­ber of ran­dom vari­ables that are under con­sid­er­a­tion — in this case the ones pro­vid­ed by the clas­si­fi­er — by allow­ing the algo­rithm to find pat­terns in the data, bring­ing to the fore­ground the prin­ci­pal con­nec­tions present. As its name indi­cates, the process is sto­chas­tic — mean­ing, one of its func­tions is ini­tial­ized ran­dom­ly. As a result, each run could out­put a unique visu­al­iza­tion.³ In Sto­chas­tic Cities, the process was run four times. Although the over­all com­po­si­tions of the result­ing images were dis­tinct, the asso­ci­a­tions between and among the com­po­nents fol­lowed a con­sis­tent log­ic, maybe sim­i­lar to the one we would obtain if we asked a human who had nev­er seen a city from above to orga­nize those 3,750 images.

The four runs gen­er­at­ed sim­i­lar effects, such as a vis­i­ble shift in scale, per­haps because the parks were not joined as one, like the water, but rather were scat­tered along the shore, as hybrid links between nat­ur­al and built envi­ron­ments. The new images of Chica­go appear to be orga­nized by typol­o­gy, pre­sent­ed as a lim­it­less city that has lost its cen­ter. A gener­ic city at first sight, a mix of pre­dictabil­i­ty and arbi­trari­ness in the treat­ment of met­ro­pol­i­tan sys­tems fos­ters uncan­ny sit­u­a­tions, such as the way piers enter the water in an accen­tu­at­ed embrace, or the con­sis­tent recon­struc­tion of McCormick Place as a vor­tex of turn­ing roads, not to men­tion the impos­si­ble con­nec­tions of diag­o­nals and curves that gen­er­ate mag­i­cal in-betweens paired with seam­less jux­ta­po­si­tions of build­ings that total­ly erase the mosa­ic qual­i­ty of the grid in the sec­tions they occupy.

What does the algo­rithm know? Not that McCormick Place is the largest con­ven­tion cen­ter in North Amer­i­ca, nor that the Chica­go Riv­er had its flow reversed. It knows only what it has been taught, and that mix of speci­fici­ty and alter­nate log­ic nur­tures an unex­pect­ed urban­ism. Back and forth, through these oper­a­tions, the mar­gin becomes blur­ry between what the com­put­er ​“sees” and what our eyes read in the out­puts. This reflex­iv­i­ty is present in the real dynam­ics of urban­ism as well as in the space of com­pu­ta­tion­al imag­i­na­tion, where the rela­tion­ship between humans and machines becomes that of col­lab­o­ra­tors in a spec­u­la­tive, gen­er­a­tive venture.

Look­ing again at the new images of Chica­go, I dis­cov­ered a choice the algo­rithm had made, one I had not noticed before and that sur­prised me. A lit­tle square of water, iso­lat­ed between roads and away from sim­i­lar units, suf­fered that same fate in each of the four iter­a­tions: a lone­ly pond that, giv­en the mag­i­cal ratio­nal­i­ty of the images, my mind could not explain. I will sure­ly nev­er know why the algo­rithm chose to place that unit in that way, but I embrace that. In remain­ing beyond the scope of my under­stand­ing, the mys­ter­ies of the algo­rithm make evi­dent the col­lab­o­ra­tive nature of our shared work.