Page 298 - The Black Box Book

C U R AT O R O F TA S T E
In a recent experiment, we demonstrated that neural networks can detect
and categorize artistic techniques. However, could this technology be uti-
lized to classify images into less recognizable categories? For example, to
estimate its association with a particular art school, group of artists, exhi-
bition, show, or individual collection? Is it possible to go even further and
create an algorithm capable of replicating curatorial taste? If today’s gen-
erative neural networks produce music, writings, or paintings based on
databases of works by long-dead artists, why couldn’t we also resurrect
legendary collectors like Lorenzo de’ Medici or Daniel-Henry Kahnweiler
and ask them whether they would select this or that piece to their collec-
tion? As an alternative, how about looking for paintings that are formally
comparable to these collections and determining whether these similarities
point to any deeper connections between the pieces of art?
There is a tendency for major styles as well as some smaller ensem-
bles of works to have a similar aesthetic. When we speak of workshops,
schools or circuits in the visual arts, we frequently refer to formal simi-
larities between individual artifacts. Other groups, however, were formed
through a variety of external circumstances, and their composition can
be visually quite disparate. Artists were admitted to aristocratic courts
on the basis of recommendations from masters or professors; they joined
societies or groups on the basis of shared opinions and attitudes; they
may have been included in an exhibition through the efforts of gallerists
or art dealers, and so forth. The machine vision algorithm cannot take
into consideration these (not only) social settings because they are not
inherent in the visible layer of the artwork. The absence of context can
be both advantageous and disadvantageous, on the one hand, historical
knowledge is omitted, but on the other hand, the works can be compared
with the “innocent eye” of a program unburdened by a predefined context.
As foreshadowed in the introduction, this experiment focuses on
the tastes recorded in the collections of Emperor Rudolf II. These col-
lections were intended to encompass the total diversity of the world at
the turn of the 16th and 17th centuries and to be a representative manifes-
tation of the metaphorical concept of the Theatrum Mundi. The collections
included objects from every known continent, representing a wide range
of human interests and diverse components of animate and inanimate
nature. They included minerals, semi-precious and precious stones,
fossils of plants and animals, stuffed animals, skeletons, skins, antlers,
bezoars, technical devices, clocks, astronomical instruments, historical
artifacts, copies of ancient works, and contemporary (late Renaissance and
Mannerist) works of art in the fields of glass, jewelry, sculpture, drawing,
printmaking, and painting.
From the perspective of this experiment, the most noteworthy is
Rudolf’s painting collection, which represents the source of paintings for
the sample dataset. This collection included allegorical and mythological
works, genre paintings, especially landscapes, still lifes with flowers and
298
various representations of fauna and flora. It also included a set of por-
traits and allegorical works reinforcing the emperor’s status, as well as
a number of paintings with erotic overtones. On the other hand, depictions
of religious motifs were not often found in the imperial halls, although the
“Rudolf artists” themselves also executed a number of commissions for
church clients (Horská, 2018, p. 22).
The painters (and artists in general) who worked for Rudolf II were
seasoned masters from many artistic centers in Europe who settled in
Prague. The nature of art at Rudolf’s court was therefore initially very di-
verse. Thus, It is therefore only conceivable to speak of the “Rudolphine
style” in the context of the period at the turn of the 16th and 17th centu-
ries, when artists had been working in one place for a long time and were
influencing each other (Zlatohlávek, 2012).
The sample dataset of Rudolphine works was derived from Thomas
DaCosta Kaufmann’s book The School of Prague: Painting at the Court
of Rudolf II, which is based on contemporary inventories and archival
sources and is the most exhaustive source on painting collections of
the time. The resulting database contained 328 paintings from 29 artists.
The range of artists included well-known names such as Hans von Aachen,
Giuseppe Arcimboldo, Bartolomeo Spranger and Joseph Heintz, as well
as lesser-known artists such as Daniel Fröschl, Paul Vredemann de Vries
and Johann Geminger.
If we want to train a neural network to recognize a certain phenom-
enon, we must frequently prepare not just an example dataset containing
images of the phenomenon but also a negative set in which the observed
element does not exist. For example, if the goal is to create an algorithm
that recognizes a cat, we need to collect and label photographs that
include the animal, as well as those that do not include the animal at
all. What exactly should a picture with a “non-cat” look like? Simply put,
there should be a wide variety of photographs showing potential settings
in which the cat might be present but is not. Adopting a similar approach,
we also require pieces by “non-Rudolphines,” or painters unrelated to
the Prague court, in order to develop a program that reflects the aesthetic
preferences of the Rudolphine painting collections.
How to define these “non-Rudolfers”? If the taste of the imperial
curators is a difficult-to-grasp abstract phenomenon, how can one pick
works that do not include this phenomenon without committing biases
based on one’s own preconceptions and preferences? For the time being,
it seems most prudent to identify as non-Rudolphine all works that were
made before Rudolf’s adulthood and hence could not have been influenced
by his collecting tastes at the time they were created. However, the nega-
tive dataset did not include any works produced after the Emperor’s death.
The rationale behind this is the belief that the themes, style, and taste
of Rudolphine painters persisted long after the monarch’s death. Indeed,
finding works that remained influential subsequent to the Monarch’s reign
was one of the aims of this experiment.
299