Google Disrupts Curating Via Artificial Intelligence
In artnet’s predictions for 2017, I wrote that it would be the year Artificial Intelligence would finally crack the problem of curating. That was meant as a joke—but it turns out that I was already behind the times. The whiz kids at Google’s non-profit cultural arm, Google Cultural Institute, have spent the year trying to imagine just that.
Head to the “Experiments” section of the Cultural Institute website and you will find a catalogue of how they have been attempting to apply “machine learning” to the question of organizing artworks. Perhaps the experiment that best showcases their particular brand of cultural gimcrackery is called “X Degrees of Separation.”
Google Cultural Institute has secured millions of high-quality images of artworks and artifacts from hundreds of partner museums around the world. “X Degrees” lets you pick any two images from this library. Its algorithm then conjures up a series of steps that connect the two images visually, using other artworks from the trove.
The result looks like a logical evolution, if you squint. The connections are conjured out of a machine’s understanding of artistic similarities. Half the pleasure is the unexpected leaps this produces.
Thus, when I select Raphael’s Madonna in the Meadow (from the Kunsthistorisches Museum in Vienna) and a photo of a 2013 Nick Cave sound suit (from the Baltimore Museum of Art), “X Degrees” creates a visual pathway that passes through a Giulio Romano Madonna and Child, to a Cranach Pieta, to an Eliot Porter picture of crab legs, and a Wangechi Mutu painting of a bust composed of amoeboid, collage-like forms.
Another recent Experiment, called “Tags,” scans the trove, then generates a wave of keywords “without the intervention of humans” that “reflect how a computer ‘sees’ Artworks.”
Apparently, a computer “sees” artworks according to categories like “vertebrate” and “hairstyle,” as well as genres like “monochrome photography” and “classical sculpture.” The Tags include the eccentrically specific, like “boats and boating—equipment and supplies,” and the unexpectedly abstract, like “material property,” as well as curveballs like “massively multiplayer online role-playing games.”
That last category, incidentally, does not yield up the expected images of World of Warcraft and EVE Online. Instead, it gives you a diorama of Cagayan Valley Life circa 750,000 BC from the Ayala Museum in the Philippines and a wallpaper depicting “Views of the American War of Independence” by the firm Zuber & Cie, from the Cooper Hewitt collection.
Such quirks, I guess, are part of the fun of learning how a computer “sees” art.
A few other Experiments involve organizing the Google trove of art images into 3-D landscapes. “t-SNE Map,” for instance, gives you a view of a landscape of rolling hills, composed of points. Zoom in on the terrain, and you discover that the points are actually images of artworks. The topography is formed by how the computer has decided to sort and cluster artworks in relation to one another, based on its understanding of aesthetic similarities.
“The algorithms only ‘looked’ at the artworks,” the description explains. “No meta data was used, the visual similarity was calculated with a computer image algorithm used in Google Search purely based on the images.”
Still another Experiment, called the “Curator Table,” claims inspiration from “the principle of laying out prints on a table when planning an exhibition.” It conjures a 3-D image landscape that you can sort by artist, time, or color, promising the ability to forge new connections quickly out of vast amounts of art. In a TED talk, Google Cultural Institute boss Amit Sood explained its origin:
Beyond the pretty picture, beyond the nice visualization, what is the purpose, how is this useful? This next idea comes from discussions with curators that we’ve been having at museums, who, by the way, I’ve fallen in love with, because they dedicate their whole life to try to tell these stories.
