Trevor Paglen: ‘This is Not an Apple.’ Or is it?

Trevor Paglen: From ‘Apple’ to ‘Anomaly’
The Curve, Barbican Centre, London
26.09.19 – 16.02.20

The exhibition begins, like the biblical story of mankind, with an apple. Or rather, it begins with a glut of photographs of apples: red apples, green apples, yellow apples, sliced apples, bushels of apples, an apple with the word ‘Google’ carved into it. These images are pinned to a wall and clustered around a single, simple noun: ‘apple’.

The exhibition features a project by artist Trevor Paglen, who culled more than 30,000 images from the online database ImageNet, printed them out in small-scale versions, and pinned them to a long curved wall that unfolds like a sprawling mosaic of the universe, categorised by noun: ‘apple’, ‘soil’, ‘valley’, ‘syringe’, ‘mascot’. It is breathtakingly expansive, like an endless quilt of colours and shapes, a visual catalogue of everything under the sun. Especially seen from a distance of a few feet, it looks like a kind of shimmering mirage of animals, plants, minerals and people.

This infinite visual library is possible because ImageNet is one of the largest publicly available data sets of images online. It is also the foundation of most of the world’s image recognition technology. Algorithms are trained to see the world using its data, which was amassed over the course of nearly a decade by workers who matched images to associated words for pennies on the task website Amazon Mechanical Turk. The result was a data set of 14 million images, catalogued and labelled, so that machines could learn to match images of roses to the word ‘rose’. Paglen makes visible this wild taxonomy in From ‘Apple’ to ‘Anomaly’.

Paglen, an American artist, is known for work that deals with technology, image culture, and the way networked systems of surveillance are integrated invisibly into our everyday lives. His artworks are often ambitious in both scale and methodology. He learned to scuba dive in order to photograph fibre-optic cables on the ocean floor, and launched a satellite into perpetual orbit, in what he called a ‘time capsule’ that critiqued the commercialisation of space. He flew over the headquarters of the United States National Security Agency, aerially photographing the expansive complex of buildings and land. There is an element of activism to his work: the National Security Agency, for instance, objected to Paglen’s aerial flight, but he went ahead with the project anyway. He is less didactic than cerebral, however, and his works visualise highly conceptual and often hidden structures of surveillance.

Sometimes, Paglen uses the tools of the technology he is critiquing, as he did in 2019, when he accidentally entered a new realm of internet fame. He and researcher Kate Crawford released the project ImageNet Roulette, which allowed users to upload their own photos to see how an algorithm trained on ImageNet would label them. The project went instantly viral, as users started encountering racist stereotypes. One twenty-four-year-old Black man found that the algorithm labelled his smiling selfie ‘wrongdoer’ and ‘offender’; an Asian-American woman was labelled ‘gook, slant-eye’. More and more reports of algorithmic bias rolled in, until ImageNet eventually said that it would take down 600,000 images in its systems. ImageNet Roulette was a participatory, networked artwork that immediately, shockingly, exposed the sexism, racism and other biases embedded in image recognition technologies. From ‘Apple’ to ‘Anomaly’, which was based on the same research, exposes how these biases are based in the very foundations of machine learning.

Part of being human is possessing the ability to tell apples from oranges. It’s also being able to tell a green apple from a red one, an apple tree from an apple seed, and an apple – object, fruit – from the Apple logo. Naming and organising the world into categories was one of Adam and Eve’s first tasks in Genesis. It is something we are taught to do in primary school, when we learn to organise blocks by shape and size. Our ability to recognise and place something into a category – black, white, round, square, dog, cat, wolf – is both a primal instinct and a sophisticated impulse that can lead to profiling and stereotyping.

Because of its centrality to human experience, taxonimising is also an essential part of the growing field of image recognition technology; machines need to learn how to break the world into human concepts and categories in order to do any kind of ‘recognition’. Like children, algorithms must be taught the shape and size of an apple, and what makes it distinct from an orange. But breaking the universe into categories is not always so easy, and can lead to big philosophical questions, such as ‘What makes an apple an apple?’ Paglen acknowledges this with a wink to Surrealist painter René Magritte. Near the entrance to the exhibition, Paglen included a reproduction of Magritte’s Ceci n’est pas une pomme, a painting of an apple beneath the phrase ‘This is not an apple’. Magritte was toying with the distinction between an object and the representation of it, the distance between a painted apple and an edible Granny Smith. In Paglen’s updated version, Magritte’s painting has been overlaid with an algorithm that identifies it: ‘red and green apple’. The machine has declared, with no ambiguity, that ceci est une pomme, after all.

Philosophical questions about representation and reality get flattened by algorithms that break the world into definite categories based on human assumptions. The ambition of ImageNet is to make all things knowable by splitting them into groups. This leads to some absurdity, evident in the exhibition. From ‘apple’, we move to ‘apple tree’ and ‘apple orchard’, which is simple enough. Then there is ‘valley’ and ‘soil’, the general realm of the pastoral. And then there is ‘labourer’. Here we see men at work, bent over in rice paddies and irrigated fields. Almost all are non-white. Soon, travelling along the curved wall, we arrive at ‘investor’, a group composed almost entirely of white men in suits and ties, pointing at whiteboards or bent over laptops. We begin, perhaps, to see the cracks in the foundation of this labelled visual universe. We recognise stereotypes about race, class and gender, spat back to us by machines.

The nouns Paglen chooses trend increasingly towards the abstract, and the matching becomes more and more chaotic. The word ‘segregator’ is associated with images of both Barack Obama and George W. Bush. Photos associated with ‘wine drinker’ and ‘alcoholic’ bleed together on the wall, but are subtly different: a white man sniffing wine is a ‘wine drinker’, while a Black woman with a large salted Martini rim is an ‘alcoholic’. Paglen visualises the consequences of a taxonomy that’s based on collective human biases, which are then replicated exponentially by machines. He also shows how difficult it is for machines to make sense of any kind of abstraction, or for us to try to teach them what a ‘segregator’ would look like.

One of the paradoxes is that there is quite a bit of beauty in From ‘Apple’ to ‘Anomaly’, formally and conceptually. The repetition of fried eggs and sunsets side by side leads to surprising visual resonance. Even the glitches can be serendipitously beautiful, as when a cheetah is mistakenly classified as a ‘honeycomb’. The project is strangely awe-inspiring; after all, we are in the presence of an attempt to make a catalogue of the physical world, the vast and expansive fields of human and non-human experience. ImageNet has quasi-religious dimensions, as it attempts to make the entire universe knowable by name.

But in From ‘Apple’ to ‘Anomaly’, Paglen probes the darker underside of this effort to understand by categorising. As in ImageNet Roulette, this project elucidates algorithmic bias, but it also makes clear a more foundational absurdity – the idea that we can easily break the universe down into clearly defined categories and teach machines to understand what they are. The whole system collapses by the end of From ‘Apple’ to ‘Anomaly’, and the exhibition reminds us that our vast, strange, complicated universe is fundamentally impossible to contain within easy-to-define boundaries.





作为全球最大的在线公开图像数据库之一,ImageNet让这个无限的视觉图录成为可能。它也是世界上大多数图像识别技术的资料基础。这些数据由工作人员耗费近十年时间,在任务网站“Amazon Mechanical Turk”上以几分钱的价格,将图像与相关单词进行配对后所积累起来的。其结果是打造了包含1400万张图像的庞大的数据库,系统还对其进行了分类和标记,如此一来,电脑就可以学会将玫瑰的照片与单词“rose”进行匹配。帕格伦在“从‘苹果’到‘异常’”中展示了这种疯狂的分类。


帕格伦有时会使用他所批评的技术工具,就像他在2019年所做的那样,当时他意外进入了“网红”的新领域。帕格伦和研究人员凯特·克劳福德( Kate Crawford )上线了“ImageNet轮盘赌”项目,该项目允许用户上传自己的照片,然后体验根据ImageNet训练的算法,这些照片是如何被贴上标签的。“ImageNet轮盘赌”迅速变得热门,因为用户开始遭受种族主义的刻板歧视。一名24岁的黑人男子发现,算法将他的微笑自拍标记为“罪犯”和“违法者”;一名亚裔美国女性被标记为“眯着眼睛的韩国人”。越来越多的关于算法偏见的举报涌入,直到ImageNet最终表示将删除系统中的60万张图片。“ImageNet轮盘赌”是一项具有参与性、网络化的艺术实践,它令人震惊地暴露了被默默嵌入图像识别技术中的性别歧视、种族歧视和其他偏见。“从‘苹果’到‘异常’”也基于同样的研究,艺术家揭示了这些偏见是如何在机器学习的基础上形成的。


由于以人类经验为中心,分类也是图像识别技术发展领域的重要构成要素;机器需要学习如何将世界按照人类的概念加以区分,以便进行任何类型的“识别”。如同孩子,任何算法必须牢记苹果的形状和大小,以及它与橙子的区别。但将宇宙划分为不同类别却并不总是那么容易,甚至可能遭遇一些重大的哲学问题,比如,是什么让苹果成为苹果?帕格伦向超现实主义画家勒内·马格利特( René Magritte )眨了眨眼睛,通过展览指涉了这一点。在展览入口附近,佩格伦展出了一幅马格利特名作《这不是一颗苹果》( Ceci n ‘est pas une pomme )的复制品,画中的苹果下面写着“这不是一颗苹果”。马格利特在玩味物体与它的表现形式之间的区别:一个涂了漆的苹果和一个可食用的绿苹果之间的距离。在帕格伦更新后的版本中,马格利特的画作上被覆盖了一种算法,识别出画里是“红色和绿色的苹果”。机器已经毫不含糊地宣布:无论如何,这是一颗苹果。