Trevor Paglen (Bloom)
PACE Gallery, October 2020
The most interesting and mysterious component of the exhibition consists of large dye sublimation prints of photographs of flowers and trees, but with their colours and perhaps shapes modified by the class probabilities of an image classifier applied to parts of the photograph. It is written “They do not represent real-to-life colours so much as what the AI thinks the different parts of the images are.” The work inspires the possibility of using hierarchically convolved imageNet classifiers to interfere with or (re-write) the images that it sees, e.g. semantically segmenting images.
Less interesting is the systematic representation of datasets, e.g. video datasets of drivers, again class of action coded by colour, but shown very small like film rolls spread out into a grid. Similar raw data presentation occurs with the handwriting samples for training classifiers but formed into the words of the American constitution, or twitter data with emotional valence classifications, all written out very small so it looks like a big square from a distance. This falls into the class of scientific art which consists of Figures that would not have been positively reviewed by an academic reviewer who would have suggested making the font bigger. His earlier work ‘selections from an imageNet dataset’ which was placed on an enormous wall is similar, but more magnificent, less of a failed scientific diagram.
There are also a few nice hand drawn Indian ink representations (also quite scientific diagramy) of the features used in face recognition, with little springs on lines used to represent variables, I like this as a positively reviewed scientific figure. And finally there is the image that is constructed from polygons by optimizing a class probability of a classifier trained on the faces of people Mr Paglen knew. The choice of generative mark is excellent, each polygon overlapping with transparency, producing quite a ghostly image, as if reconstructed from a memory crime mugshot of a distantly remembered face.
I shall completely ignore the live classifier video work looking at classifier biases because it ignored me as a person of colour, only recognizing my friend as platinum blond cleaner.
Paglen makes magnificent and epic images, the moral of the story is that one shouldn't be afraid to spend a lot of money on enormous high resolution dye sublimation prints. I love his more recent work e.g. Hough lines 2019, which shows lines superimposed on a photograph of clouds, as if someone has gone up to the sky and tied string to clouds, it is precarious and majestic. The shape of clouds exhibition suggests how image processing can be used to superimpose drawings on photos, something artists have been doing for a while now, e.g. sticking string stepladders to mountains.
I have tried a similar thing , where I allow a random neural network to foveate on a Magritte painting and track the vision system as it moves over the image, while it is being trained. The logical development of this process will result in a reinterpretation of the image, allowing the image to eventually be removed entirely, with only the saccadic and smooth pursuit processes remaining visible.
An important principle from Paglen is that AI can be used to show how we see an image as well as the image itself, o and also that there is a huge deal of secret and mysterious sensitivity in the production of ‘AI art’, o and that you shouldn't be afraid to print things big.