Page 274 - The Black Box Book

The last type, transformative creativity, is characterized by tran-
scending the boundaries of the previous conceptual space and developing
a new structure of thinking. For a long time, this type of creativity could
not be mastered by computers. But this changed, according to Boden,
when developers such as Karl Sims and Wiliam Latham experimented
with so-called generative algorithms in the early 1990s, which were based
on the principle of evolutionary programming and in this way generated
computer graphics of astonishing complexity (Sims, 1991, see also: Todd
& Latham, 1992; Bentley, 1999).
Meanwhile, ANNs have outdone other methods that have been
considered as a path toward the goal of creative AI in recent years and
decades. For the time being, at least. Nonetheless, the creative potentials
of ANNs are also doubted by various philosophers, artists, and humani-
ties scholars: For example, in 2016 Matteo Pasquinelli argued that ANNs
are fundamentally incapable of producing anything emphatically new.
According to him, ANNs would be able to recognize patterns in already
existing data purely inductively, but they would not be capable of what
Peirce calls the operation of abduction. Hence, ANNs would be unable to
achieve, what Margaret Boden called transformative AI.
Also, the media philosopher Dieter Mersch recently reproached
those who, like Margaret Boden, are willing to ascribe creativity to com-
puters for mostly using simplistic and historically outdated definitions of
art and culture as a basis and at the same time consistently disregarding
a specifically “epistemological dimension of the aesthetic”:
“Persistently they ignore what art makes art in the first place:
Reflexivity as the opening up of a different knowledge. Instead,
under the sign of a preference for rationalism and hard sciences,
a direct connection is drawn between ‘natural’ creative activities
such as the development of life and the ‘social’ or ‘ historical’ vir-
ulence of the arts, regardless of essential incompatibilities. If, on
the other hand, we insist on the persistence of an aesthetic reflexivity
as a criterion for art, it is because this reflexivity is not drawn from
perceptual extensions, subjective expressive excesses, or the rep-
resentation of borderline phenomena, just as little as the properties
of art can be inferred exclusively through its ‘works’; rather, these
always ‘acquire’ a transformation and displacement of art itself. Art
is always art about art; it therefore implies in every act and artifact
a transformation of the aesthetic itself, whereas most models of an
artificial creativity are oriented to continuities and remain attached
to an anachronistic nineteenth-century cult of genius, which in turn
owes itself to a vulgarization of the Kantian definition of genius”
(Mersch, 2019, p. 73).
274
Even though, we can list some examples of computer scientists who use
extremely reductionist models of creativity (see Langley et al., 1987),
still the latter accusation seems a bit sweeping. Back in the 1980s, for
example, computer scientist Marvin Minsky criticized that people tend to
exaggerate the difference between ordinary and outstanding achievements
in terms of their creative thinking (1982). At least in this case, there can
be no question of a cult of genius.
In general, a fundamental problem with the debate on AI is that
we typically value and admire those achievements of machines that we
also value and admire related to humans. If a computer beats the human
world champion in the board game Go, it is a sensation that is reported
globally in the media. However, there was much less public interest in
2017 when an AI model was able, for the first time, to recognize actions
and gestures using ANN-based methods and based on training with video
data, i.e. when it was able to master, in a basic form, what is commonly
referred to as common-sense knowledge in relation to humans (Sudmann,
2016, 2017). It was not acknowledged that the step of enabling an AI to
visually perceive and recognize its environment is much more fundamental
and far-reaching – not least for the development of advanced systems
that will eventually be capable of human-like creative behavior – than
the well-known high-profile developments like AlphaGo & Co.
GENERATIVE ADVERSARIAL NETWORKS
One may also be skeptical whether Generative Adversarial Networks
(GANs) or their more developed successors as Creative Adversarial Net-
works (CANs) will raise the performance of an AI in terms of creativity to
a new level in the medium or long term. CANs essentially copy the function-
ality of GANs but add a decisive component to the algorithmic operations:
The discriminator network still has the task of classifying images according
to whether they are real or fake, but also learns to assign these artifacts
to 25 art styles (like impressionism or surrealism, etc.). Accordingly,
the generative network still has the function to generate images that
the discriminator considers to be real, but in addition to that it is supposed
to make the assignment to these art styles impossible. The performance
of these creative algorithms thus consists in a fixed scheme of deviation
from a norm as an iterative optimization process. Hence, CANs deviate
decisively from what characterizes deviation from established art styles
with regard to human creativity, namely that as an artistic operation it
is itself always up for disposition in relation to a historical norm, i.e. it is
uncertain (Elgammal, 2017).
As a result, the computer art produced by CANs may even be more
interesting than many works by human artists, and this would be solely
275