Page 101 - The Black Box Book

DISTANT READING
Artificial neural networks are a subcategory of machine learning, one of
the most progressive disciplines in artificial intelligence today. In our case,
we use artificial neural networks as a tool of the distant reading method,
which is suitable for exploring cultural phenomena at the scale of big data.
2
This approach provides knowledge based on quantifying the occurrence
of specific parameters of the studied set of traits and their processing
with the help of statistical methods and probability theory. The outputs
of neural network processes can be used in distant reading to support
the exact interpretation of cultural phenomena through categories such
as clusters of phenomena, density or, conversely, sparsity of occurrence
within discursive formations. Moreover, they allow us to free ourselves
from the expectations, preconceptions and emotional projections that
form a natural part of our interpretation of the world and thus to view
the traces of the global pandemic from an anaesthetic distance – from
the outside, from the perspective of a non-human actor and observer.

UNSUPERVISED LEARNING
We wanted to reinforce this non-human view as much as possible
in the project, thus developing the black box motif associated with
a post-apocalyptic vision of the end of the human world and the rise
of non-human intelligence (artificial or virus intelligence?). To predict
the results, we used methods falling into the category of unsupervised
learning. This is an area of machine learning called learning without su- Figure 5: Clusters produced by applying the
pervision, which is used to search for unspecified patterns in a dataset Mean Shift algorithm to the reduced image parameters
obtained by t-SNE. Author: Štěpán Miklánek
that is not provided with predefined metadata and minimal assistance
from a programmer. In the case of the Black Box project, this involves
machine sorting of image material into groups (clustering). The neural THREE ATTEMPTS TO RID ARTIFICIAL
network was used to extract unique fingerprints from the image material;
the subsequent sorting into groups was performed using unsupervised I N TE L L IG E NC E O F T H E
learning algorithms that operate based on distance metrics (Euclidean H U M AN E XPE R I E N C E
distance and others). With this combination, we created a semi-super-
vised algorithm. The Python programming language enabled easy imple- ATTEMPT 1: THE POWER OF COLOR
mentation of available machine and deep learning modules. The input to In the first experiment, the already trained Xception network model, which
the neural network was visual material of colour images collected from is part of the freely available Keras module (Chollet and others, 2015) for
selected artists. Python, was used to extract features from the image data. (Chollet et al,
2017) In this case, it was software with trained weights on the ImageNet
dataset, one of the largest databases of manually annotated image ma-
2 Franco Moretti, who is thought to have coined the term distant read- terial, comprising more than 14 million digital images. (Deng et al., 2009)
ing, first used it in an article Conjectures on World Literature. New left The colour representations of the images, given by the R, G, and B chan-
Review, January/February 2000. Available online: https://newleftreview. nels (red, green, blue), were chosen as the primary criterion for classifying
org/issues/ii1/articles/franco-moretti-conjectures-on-world-literature
(cit. 19. 12. 2022.) the dataset’s contents. The features representing the unique fingerprint
of each image were extracted using a neural network. These features
were then converted into numerical vectors. The next step was to reduce
the number of dimensions of the vectors to just two dimensions, i.e., two
numerical values representing the location of the corresponding image
in space. For this purpose, the t-SNE (T-distributed Stochastic Neighbor
100 101