Hi, every­one. Glad to be back.

23andMe is a US-based com­pa­ny that will, for a fee, ana­lyze your genet­ic ances­try. Provide some sali­va, and they’ll send you a sum­ma­ry that com­pares your DNA with thirty-one iden­ti­fied pop­u­la­tions across the world. It seems that this sort of genet­ic analy­sis is the next big thing in fam­i­ly his­to­ry, and you may have noticed that Ancestry offers a sim­i­lar ser­vice. 23andMe, how­ev­er, also pro­vides an appli­ca­tion pro­gram­ming inter­face, an API, so that devel­op­ers can cre­ate cool third-party apps with your DNA. Don’t wor­ry, you do actu­al­ly have to pro­vide per­mis­sion. So Facebook won’t auto­mat­i­cal­ly start send­ing you friend requests based on your genes. At least not yet.

But it did­n’t take long for the eth­i­cal bound­aries of this sort of ser­vice to be test­ed. One devel­op­er cre­at­ed a genet­ic access con­trol authen­ti­ca­tion sys­tem. Using it, online sites or ser­vices could restrict access to peo­ple who had a spe­cif­ic genet­ic make­up. This was an actu­al exam­ple from the guy who devel­oped it:

Screenshot of an error message stating "Authorization Status: Invalid! You are 22% of the permitted European ancestry."

The devel­op­er’s API access was quick­ly revoked, with 23andMe not­ing that their Terms of Use pro­hib­it appli­ca­tions that con­tain, dis­play, or pro­mote hate.

Four years ago, I stood on this stage describ­ing a project that I was work­ing on with Kate Bagnall called Invisible Australians. We were and are try­ing to encour­age use of the National Archive of Australia’s col­lec­tion of records that doc­u­ment the work­ings of the White Australia Policy in quite con­fronting detail.

As an exper­i­ment, I down­loaded thou­sands of images from the Archive’s col­lec­tion data­base. Most were cer­tifi­cates used in the con­trol of non-white immi­gra­tion, visu­al­ly com­pelling doc­u­ments that include both por­trait pho­tographs and handprints. 

I ran a facial detec­tion script over these images to extract the por­traits, and cre­at­ed an online resource called The Real Face of White Australia

For a week­end project, it’s had a sig­nif­i­cant impact in the dig­i­tal human­i­ties world. Some peo­ple crit­i­cize this, though, because they thought we were actu­al­ly select­ing records based on race. This was just a mis­un­der­stand­ing of both the records and the tech­nol­o­gy that we were using. Even if I’d want­ed to, I would­n’t have had a clue back then about how I would cat­e­go­rize these por­traits by race.

Now I do.

I can sign up for an account with a ser­vice like Face++ and use their API to ana­lyze an image of a per­son­’s face to deter­mine both race and gen­der. After a bit of a break, because of life, Kate and I are get­ting back to Invisible Australians. In 2011, I had about 12,000 images from one series in the National Archives. I’ve now host­ed more than 160,000 images from twenty-two dif­fer­ent series.

But the inter­ven­ing years have also brought changes in the Australian gov­ern­men­t’s treat­ment of asy­lum seek­ers. It’s brought changes in greater pow­er for secu­ri­ty agen­cies and the nor­mal­iza­tion of elec­tron­ic sur­veil­lance. When the White Australia Policy was imple­ment­ed, por­trait pho­tographs and fin­ger­prints were the lat­est in crime-fighting tech­nol­o­gy. Just over a month ago, the Australian gov­ern­ment announced its newest nation­al secu­ri­ty weapon, a nation­al facial recog­ni­tion sys­tem to be known hence­forth as The Capability. 

It’s true, it’s true. I sus­pect they already have the movie rights in mind.

The sys­tem would assist author­i­ties in putting a name to the face of ter­ror sus­pects, mur­der­ers, and rob­bers. Tools that help iden­ti­fy faces can offer pow­er­ful new means of dis­cov­ery and analy­sis with­in the hold­ings of our cul­tur­al col­lec­tions. But can those of us who work with these tools avoid engag­ing with broad­er sys­tems of sur­veil­lance, cat­e­go­riza­tion, and con­trol? For Kate and me, the par­al­lels are just too strong. History is not just about the past.

I’m talk­ing today about two relat­ed tech­nolo­gies, facial detec­tion, and facial recog­ni­tion. Facial detec­tion sim­ply tells you if there’s a face in an image. It’s a the tech­nol­o­gy that draws lit­tle box­es around faces when you’re tak­ing pho­tos, and it’s pret­ty effi­cient and well-established. This was the tech­nol­o­gy that I used to cre­ate our wall of faces. 

Basic detec­tion is now being sup­ple­ment­ed by algo­rithms that exam­ine the shape of facial fea­tures, and the char­ac­ter­is­tics of things like skin so that they can esti­mate gen­der, age, and race. They can also tell you whether the per­son is wear­ing glass­es, and the qual­i­ty of their smile.

Facial recog­ni­tion, on the oth­er hand, first detects faces with­in an image, and then search­es for those faces with­in an exist­ing set of previously-identified images. The Capability, for exam­ple, plans to take an image and look for match­es across a series of inter­linked data­bas­es such as pass­ports and dri­ving licens­es. It’s also of course what Facebook does when it tags peo­ple in the pho­tos that you share. 

Facial recog­ni­tion is a lot trick­i­er than detec­tion, but last year Facebook announced that its DeepFace sys­tem (Don’t you love all these names?) had reached a lev­el of accu­ra­cy sim­i­lar to humans. Not to be out­done of course, Google claimed it’s FaceNet tech­nol­o­gy had pushed the bar even high­er, report­ing an accu­ra­cy of over 99%. How this trans­lates to real-world appli­ca­tions such as The Capability is not clear, but I think we can safe­ly assume that the secu­ri­ty agen­cies are head­ing down the same path.

Machines have strug­gled to match humans in find­ing and rec­og­niz­ing faces because faces are so impor­tant in sim­ply being human. Faces con­nect us to our social world. But as cul­tur­al insti­tu­tions know, faces can also con­nect us through time. Looking into the eyes of a per­son, no mat­ter how far removed through time, his­to­ry, or cul­ture, affects us. We do not mere­ly see, we feel. And this I think is where The Real Face of White Australia gains its emo­tion­al power.

Last year, in anoth­er exper­i­ment, I start­ed extract­ing faces from Trove’s mil­lions of dig­i­tized news­pa­per arti­cles. The images are of a much low­er qual­i­ty than the pho­tographs of course, but still…there’s that feel­ing of connection.

A grid of many faces cropped from old photographs

Facial detec­tion is one exam­ple of a broad­er class of com­put­er vision oper­a­tions known as fea­ture detec­tion.” You can train your com­put­er to find all man­ner of pat­terns and shapes in your images. This includes things like cats and bananas, as well as the com­po­nents of the face, eyes, nose, and mouth. So of course, one night I won­dered what would a news­pa­per dis­cov­ery inter­face based on eyes look like? 

A collection of images of eyes, some expanded to show the full face of the person

Eyes on the past has been var­i­ous­ly described as both beau­ti­ful and creepy, some­thing of which I’m rather proud. It was anoth­er week­end project, an exper­i­men­tal inter­ven­tion rather than a prac­ti­cal tool. By click­ing on eyes and faces, you can find your way to news­pa­per arti­cles, but that’s not real­ly the point. I was hop­ing to say some­thing about the fragili­ty of our con­nec­tion to the past. We glimpse past lives through tiny cracks in the walls of time. These moments may be fleet­ing, but they can also be full of meaning. 

So I kept har­vest­ing faces, and I’ve now got about 6,000 from the news­pa­pers from the 1880s through to 1913. And if you’d like to play the full dataset is avail­able for down­load from the data-sharing site Figshare.

I also built my own face API to encour­age fur­ther exper­i­men­ta­tion. While ser­vices like Face++ have APIs that take your face and pull it apart, mine just gives you ran­dom faces from the past. That’s all.

Most recent­ly, I used my col­lec­tion of faces to cre­ate a Twitter bot called The Vintage Face Depot. Tweet a pic­ture of your­self to the bot, and it will send you back a new ver­sion of your­self in which your face is over­laid with a ran­dom vis­age from my exten­sive range of vin­tage faces.

Tweaking the trans­paren­cy means that the face start to blend. You are nei­ther your­self nor them, but some­one new. Each face replace­ment also comes gift-wrapped with a link to the orig­i­nal news­pa­per arti­cle. The Vintage Face Depot tells you noth­ing new about your­self. I built it about the same time as Microsoft launched their How Old bot that uses machine learn­ing to esti­mate your age. Face Depot does noth­ing clever and yet some­times the results are uncan­ny, even unset­tling. Microsoft may be able to tell you how old you are, but Face Depot asks who you are, and push­es you in the direc­tion of a past life linked mere­ly through chance.

Of course, the next obvi­ous step is to feed the results of the Vintage Face Depot to the How Old bot, or indeed to Face++‘s API:

In a sim­i­lar vein, the devel­op­er Kurt Kaiser has been pit­ting neur­al net­work against neur­al net­work, alter­ing images of him­self using Google’s Deep Dream and upload­ing them to Facebook for DeepFace to ana­lyze and tag.

Digital artists like Adam Harvey have reverse-engineered facial detec­tion algo­rithms to devise anti-surveillance fash­ion styles. He shows how you can use make-up to dis­rupt key regions of your face such as where your nose and eyes inter­sect, and effec­tive­ly ren­der your face invis­i­ble. From face, to anti-face.

While none of these inter­ven­tion pro­vide a detailed cri­tique of state sur­veil­lance, they do high­light the con­struct­ed nature of these tech­nolo­gies. By play­ing around with their para­me­ters, we under­stand bet­ter how they work. (And that last sen­tence was mod­i­fied to be more family-friendly.)

But what’s the role of cul­tur­al her­itage orga­ni­za­tions in all of this? Libraries are already lead­ing the way in sup­port­ing online pri­va­cy. But leav­ing aside the whole liv­ing in a sur­veil­lance state thing for a moment, these tech­nolo­gies don’t just find faces, they reduce us to a set of exter­nal char­ac­ter­is­tics. We become what they can measure. 

Researchers are cur­rent­ly inves­ti­gat­ing how facial detec­tion sys­tems can be used to iden­ti­fy depres­sion. The aims are wor­thy, of course, but it’s not hard to imag­ine how, like 23andMe, such sys­tems could be used to dis­crim­i­nate rather than sup­port. Other stud­ies have explored whether human observers can tell if you’re gay or prone to infi­deli­ty by look­ing at your face. Anyone remem­ber phrenology?

With mea­sure­ment comes the pow­er to cat­e­go­rize and con­trol. These are tech­nolo­gies that enable us to be judged at a dis­tance, to be iden­ti­fied as a threat or a sales oppor­tu­ni­ty just by the way we look. Facial recog­ni­tion takes this fur­ther. Not only can we be reduced to a set of externally-verifiable mea­sure­ments, but these mea­sure­ments are assumed to some­how con­sti­tute our identity.

So run­ning Face++‘s API across a large pho­to­graph­ic col­lec­tion to iden­ti­fy, for exam­ple, pic­tures of women, seems like it could be a real­ly use­ful thing to do. But we also know that the male/female bina­ry is hope­less­ly inad­e­quate in describ­ing who we are. And iden­ti­fiers are not the same things as identities.

Cultural her­itage data is glo­ri­ous­ly messy. Even as we try and wran­gle it to fit our sys­tems, we rec­og­nize the resis­tance as some­thing pro­found­ly human. Against the pow­er of sur­veil­lance, both for secu­ri­ty and for sales, we have the oppor­tu­ni­ty, the oblig­a­tion to cel­e­brate this com­plex­i­ty, to deny the mean­ing of mea­sure­ment. You can­not know me from my face. My iden­ti­fy can not be cap­tured in your database.

Let’s use Facial detec­tion to enrich our meta­da­ta, but let’s also work with artists, devel­op­ers, and activists to chal­lenge the tech­nol­o­gy’s embed­ded assump­tions about the per­fect face.

Four years ago, I showed you our wall of faces. A few weeks ago I took those 7,000 pho­tos and ran them through a pro­gram that aver­ages facial fea­tures. I expect­ed to be crit­i­cal. I expect­ed to be annoyed. But instead of see­ing some algorithmically-generated non­sense, I just saw a person.

And there’s pow­er in that. 


Further Reference

Tim’s own post­ed tran­script of this pre­sen­ta­tion. (Not dis­cov­ered until near­ly com­plet­ed here.)