麻豆传媒

A COUPLE of days before the opening of Trevor Paglen鈥檚 latest photographic installation, From 鈥淎pple鈥� to 鈥淎nomaly鈥�, a related project by the artist all over the papers.

ImageNet Roulette is an online . The website invites you to provide an image of your face. An algorithm will then compare your face against a database called ImageNet and assign you to one or two of its 21,000 categories.

ImageNet has become one of the most influential visual data sets in the fields of deep learning and AI. Its creators at Stanford, Princeton and other US universities harvested more than 14 million photographs from photo upload sites and other internet sources, then had them manually categorised by some 25,000 workers on Amazon鈥檚 crowdsourcing labour site Mechanical Turk. ImageNet is widely used as a training data set for image-based AI systems and is the secret sauce within many key applications, from phone filters to medical imaging, biometrics and autonomous cars.

According to ImageNet Roulette, I look like a 鈥減olitical scientist鈥� and a 鈥渉istorian鈥�. Both descriptions are sort-of-accurate and highly flattering. I was impressed. Mind you, I鈥檓 a white man. We are all over the internet, and the neural net had plenty of 鈥渕y sort鈥� to go on.

Spare a thought for Guardian journalist Julia Carrie Wong, however. she was a 鈥済ook鈥� and a 鈥渟lant-eye鈥�. In its attempt to identify Wong鈥檚 鈥渟ort鈥�, ImageNet Roulette had innocently turned up some racist labels.

From 鈥淎pple鈥� to 鈥淎nomaly鈥� also takes ImageNet to task. Paglen took a selection of 35,000 photos from ImageNet鈥檚 archive, printed them out and stuck them to the wall of the Curve gallery at the Barbican in London in a 50-metre-long collage.

The entry point is images labelled 鈥渁pple鈥� 鈥� a category that, unsurprisingly, yields mostly pictures of apples 鈥� but the piece then works through increasingly abstract and controversial categories such as 鈥渟ister鈥� and 鈥渞acist鈥�. (Among the 鈥渞acists鈥� are Roger Moore and Barack Obama; my guess is that being over-represented in a data set carries its own set of risks.) Paglen explains: 鈥淲e can all look at an apple and call it by its name. An apple is an apple. But what about a noun like 鈥榮ister鈥�, which is a relational concept? What might seem like a simple idea 鈥� categorising objects or naming pictures 鈥� quickly becomes a process of judgement.鈥�

The final category in the show is 鈥渁nomaly鈥�. There is, of course, no such thing as an anomaly in nature. Anomalies are simply things that don鈥檛 conform to the classification systems we set up.

Halfway along the vast, gallery-spanning collage of photographs, the slew of predominantly natural and environmental images peters out, replaced by human faces. Discrete labels here and there indicate which of ImageNet鈥檚 categories are being illustrated. At one point of transition, the group labelled 鈥渂ottom feeder鈥� consists entirely of headshots of media figures 鈥� there isn鈥檛 one aquatic creature in evidence.

Scanning From 鈥淎pple鈥� to 鈥淎nomaly鈥� gives gallery-goers many such unexpected, disconcerting insights into the way language parcels up the world. Sometimes, these threaten to undermine the piece itself. Passing seamlessly from 鈥渁ndroid鈥� to 鈥渕inibar鈥�, one might suppose that we are passing from category to category according to the logic of a visual algorithm. After all, a metal man and a minibar are not so dissimilar. At other times 鈥� crossing from 鈥渃offee鈥� to 鈥減oultry鈥�, for example 鈥� the division between categories is sharp, leaving me unsure how we moved from one to another, and whose decision it was. Was some algorithm making an obscure connection between hens and beans?

Well, no: the categories were chosen and arranged by Paglen. Only the choice of images within each category was made by a trained neural network.

This set me wondering whether the ImageNet data set wasn鈥檛 simply being used as a foil for Paglen鈥檚 sense of mischief. Why else would a cheerleader dominate the 鈥渟aboteur鈥� category? And do all 鈥渄ivorce lawyers鈥� really wear red ties?

This is a problem for art built around artificial intelligence: it can be hard to tell where the algorithm ends and the artist begins. Mind you, you could say the same about the entire AI field. 鈥淎 lot of the ideology around AI, and what people imagine it can do, has to do with that simple word 鈥榠ntelligence鈥�,鈥� says Paglen, a US artist now based in Berlin, whose interest in computer vision and surveillance culture sprung from his academic career as a geographer. 鈥淚ntelligence is the wrong metaphor for what we鈥檝e built, but it鈥檚 one we鈥檝e inherited from the 1960s.鈥�

鈥淭he group labelled 鈥榖ottom feeder鈥� consistsentirely of headshots, there isn鈥檛 one aquatic creature in evidence鈥�

Paglen fears the way the word intelligence implies some kind of superhuman agency and infallibility to what are in essence giant statistical engines. 鈥淭his is terribly dangerous,鈥� he says, 鈥渁nd also very convenient for people trying to raise money to build all sorts of shoddy, ill-advised applications with it.鈥�

Asked what concerns him more, intelligent machines or the people who use them, Paglen answers: 鈥淚 worry about the people who make money from them. Artificial intelligence is not about making computers smart. It鈥檚 about extracting value from data, from images, from patterns of life. The point is not seeing. The point is to make money or to amplify power.鈥�

It is a point by no means lost on a creator of ImageNet itself, Fei-Fei Li at Stanford University in California, who, when I spoke to Paglen, was in London to celebrate ImageNet鈥檚 10th birthday . Far from being the face of predatory surveillance capitalism, Li leads . Wong, incidentally, won鈥檛 get that racist slur again, following that it was removing more than half of the 1.2 million pictures of people in its collection.

Paglen is sympathetic to the challenge Li faces. 鈥淲e鈥檙e not normally aware of the very narrow parameters that are built into computer vision and artificial intelligence systems,鈥� he says. His job as artist-cum-investigative reporter is, he says, to help reveal the failures and biases and forms of politics built into such systems.

Some might feel that such work feeds an easy and unexamined public paranoia. Peter Skomoroch, former principal data scientist at LinkedIn, thinks so. He calls ImageNet Roulette junk science, and : 鈥淚ntentionally building a broken demo that gives bad results for shock value reminds me of Edison鈥檚 war of the currents.鈥�

Paglen believes, on the contrary, that we have a long way to go before we are paranoid enough about the world we are creating.

Fifty years ago it was very difficult for marketing companies to get information about what kind of television shows you watched, what kinds of drinking habits you might have or how you drove your car. Now giant companies are trying to extract value from that information. 鈥淚 think,鈥� says Paglen, 鈥渢hat we鈥檙e going through something akin to England and Wales鈥檚 Inclosure Acts, when what had been de facto public spaces were fenced off by the state and by capital.鈥�