Patty deGrandpre, Progression (2018)
- Engraving with paper dots (unique triptych), 3 x 8 inches each. . Printed and published by the artist, Beverly, MA.. .
This iteration of the Art in Print Prix de Print has been judged by a custom designed AI system. The Prix de Print is a bi-monthly competition, open to all subscribers, in which a single work is selected by an outside juror to be the subject of a brief essay.
In keeping with the theme of this issue of Art in Print, the latest winner of the Prix de Print has been selected by a machine learning system. Over the past eight years, machine learning—a subset of artificial intelligence research in which algorithms “train” by examining data and observing patterns (rather than receiving detailed instruction from human pro-grammers)—has produced remarkable results. These accomplishments range from conquering complex games such as Go to issuing medical diagnoses, to improving linguistic translation. Often well-trained algorithms can spot patterns that humans fail to see, a distinct advantage over the old “hard-coded” AI systems, whose capabilities were limited to the expertise of their creators.
The fields of vision and aesthetics have traditionally offered a particular challenge to machines. For years, programmers struggled to articulate in code visual classifications that humans find trivially easy: something as simple as constructing rules that can identify whether a handwritten digit in a photo is a nine or an eight, for instance, proves incredibly challenging. Distinguishing between more complex objects, such as chairs, buildings and pets, is harder still. Machine learning systems, on the other hand, thrive in settings where knowledge is hard to codify and data is abundant: neural networks (a powerful AI architecture that has become popular in recent years) can look at many photos of cats and dogs and eventually learn to distinguish the two classes. For the Prix de Print, we worked with AI researchers to train a network1 similar to those used for differentiating digits or animals, but instead of using symbols or mammal images, we used the full history of all preceding Prix de Print competitions and the decisions of the previous (human) jurors as training data. The network’s task was to seek patterns in what makes a print compelling to human viewers.2 Recognizing that quality in a work of art is a product of both content and style, we encouraged the system to be alert to this distinction, by altering the visual styles of some of the training data images while keeping the image content the same (taking a “good” image and making it “worse.”)
It must be acknowledged, however, that given the comparatively small data set and the subjectivity and contextual complexities of assessing artistic merit, the output should not be taken as scientifically meaningful or critically definitive. One might think of it as a conceptual and recreational exercise rather than a replacement for other modes of judgment.
In the end, the system selected Patty deGrandpre’s engraving triptych Progression (2018) as the winning entry, and did so by quite a wide margin. A human, asked to emulate previous decision-making might well have focused on a categorizing quality like figuration, since the majority of previous winning entries have been figurative. Progression, however, is a geometrical abstraction. Has the system deduced a subtler set of aesthetic criteria, or was it responding to something simplistic like color preference? We do not know. What is certain is that, to human eyes, it zeroed in on a work that rewards our visual and conceptual attention.
- The most successful and widely used networks for handling visual data are convolutional neural networks (CNNs), introduced in 1998 by Yann LeCun. In recent years, many scientists have made valuable contributions to improving the performance of CNNs. For this contest we used VGGnet, a convolutional architecture designed by Karen Simonyan and Andrew Zisserman that achieved popularity thanks to its strong performance in the 2014 ImageNet competition.
- The data set included successful and unsuccessful Prix de Print entries, as well as other contemporary prints reviewed and discussed in the journal. Even state-of-the-art models are much less data-efficient than humans (requiring thousands of images of cats, for instance, in order to identify a cat as well as a human would after observing only one). Having a large volume of images enables the system to pick up on common visual elements, such as gesture, line, chromatic shifts, etc. We further augmented the training set with modified versions, as explained above, to generate crisper decision boundaries. We also used parameter values in the lower network layers from a network that had previously learned to identify a set of entities from ImageNet, a canonical open-source training dataset that includes several thousand categories of objects. This is because many of the more fundamental visual “primitives” (basic building block elements) such as curves and edges, are shared across these classes and can be learned more effectively by observing a large number of diverse entities than by focusing exclusively on prints (somewhat similar to how humans’ visual assessments of art are informed by visual knowledge accumulated from years of observations of the surrounding world).