Page 269 - The Black Box Book

by Michael Noll in his 1967 essay The Digital Computer as a Creative EDMOND DE BELAMY, PRELUDE
Medium. There he says:
TO THE BOOM IN AI CREATIVITY
“An experiment was performed using Piet Mondrian’s ‘Composition The renewed interest in the connection between art, creativity and AI
With Lines’ (1917) and a computer-generated picture composed can be related to one specific recent event. On October 25, 2018, New
of pseudo-random elements but similar in overall composition to York Christie’s made headlines with the news that for the first time in
the Mondrian painting. Although Mondrian apparently placed the history of the renowned auction house it has been auctioned a paint-
the vertical and horizontal bars in his painting in a careful and orderly ing produced by an algorithm (Cohn, 2018). And an anonymous bidder
manner, the bars in the computer-generated picture were placed ac- had succeeded to purchase it over telephone for $432,500 which was
cording to a pseudorandom number generator with statistics chosen 40 times more than originally expected.
to approximate the bar density, lengths, and widths in the Mondrian The image, which was supposedly generated by an AI, is hardly
painting. Xerographic copies of the two pictures were presented, side spectacular. It shows the blurred portrait of a fictitious person named
by side, to 100 subjects with educations ranging from high school Edmond de Belamy who, in view of his clothing, could perhaps be mis-
to postdoctoral; the subjects represented a reasonably good sam- taken for a French clergyman of the 18 century. But this association
th
pling of the population at a large scientific research laboratory. They is literally in the eye of the beholder, if one considers that the AI model
were asked which picture they preferred and also which picture of was trained with 15.000 images from different periods of art history be-
the pair they thought was produced by Mondrian. Fifty-nine percent of fore generating this portrait, which is why iconographic efforts of image
the subjects preferred the computer-generated picture; only analysis quickly reach their limits. And instead of an artist’s signature,
28 percent were able to identify correctly the picture produced by there is at the lower right edge of the picture a section of the algo-
Mondrian. In general, these people seemed to associate the random- rithm (min G max D Ex[log(D(x))]+Ez[log(1-D(G(z))]) with the help of which
ness of the computer-generated picture with human creativity whereas the image was generated. Obviously, this reference is primarily a nice PR
the orderly bar placement of the Mondrian painting seemed to stunt.
them machine-like. This finding does not, of course, detract from But the Parisian artist group Obvious, consisting of the members
Mondrian’s artistic abilities. His painting was, after all, the inspiration Hugo Caselles-Dupré, Pierre Fautrel and Gauthier Vernier, responsible
for the algorithms used to produce the computer-generated picture, for the series of generative pictures La Famille de Belamy, including
and since computers were nonexistent 50 years ago, Mondrian could the Edmond de Belamy picture, is unlikely to share this view. Instead of
not have had a computer at his disposal.” (Noll, 1967, p. 92). merely creating a work of art with AI, the project’s ambition, according
to Obvious, was nothing less than to “explain” and “democratize” AI
As the media scholar Jens Schröter rightly pointed out, at that time, with art (Vincent, 2018, online). The latter, however, obviously fits neither
Noll was particularly concerned with finding algorithms that would enable the framework of the auction nor the history of the portrait’s creation.
the production of a work of art that could be recognized as such (Schröter, The Parisian artists used, or more precisely further developed, a learning
2019, p. 303). This approach can be applied to the information-aesthetic algorithm that was actually put online as an open-source software by
experiments of the mathematician Frieder Nake, thus it easily exposes Robbie Barrat, a young man from West Virginia. The artwork was gen-
itself to the accusation that it reduces art to a purely ahistorical and for- erated using so-called Generative Adversarial Networks (GANs), which
malistic enterprise. But according to Schröter, the works of Noll and other were in turn originally developed, among others, by the computer scientist
representatives of information aesthetics could very well be understood Ian Goodfellow.
as historical reactions to the potentials of art in general, in this case GANs are a variant of ANNs in which a generator network and
computer-produced art (Schröter, 2019, p. 304). a discriminator network compete against each other in a zero-sum game.
Following the training phase, one (generative) network has the function
of generating sample data that could plausibly originate from the original
data set, although they were only generated artificially, i.e. they are ‘fake’.
The other discriminative network in turn has the task to classify the gen-
erated data according to whether they originate from the original training
data set, i.e. whether they are ‘real’ or not. The training of the GANs is
repeated until the generative network produces sample data that is pre-
dominantly accepted by the discriminative network as ‘real’.

268 269