Page 104 - The Black Box Book

ATTEMPT 2: COMMON SENSE
Considering the results of the first experiment, a method of content
analysis of the dataset was chosen, which suppresses the value of
the colour of the images by converting them into grayscale. Thus,
the images were represented by only one channel, with the values of each
pixel indicating the grayscale. The features were extracted from the images
and thus modified in the same way as in the first experiment, and again
dimensionality reduction followed. The goal was to program the software
to classify the dataset into unspecified categories automatically. For
this purpose, several commonly used clustering algorithms were tested,
which work on the principle of comparing distances between points in
the plane, focusing on finding separate clusters within the spatial projection of
the dataset. Figure 7 shows the output of the Mean Shift algorithm
(Jin et al., 2017, p. 806–808), which we evaluated as the most suitable
for our purposes.
In this case, the artificial neural networks classified the dataset
according to thematic headings with greater or lesser accuracy across
the supplied digital image sets. However, we were still working with neu-
Figure 7: Thematic clusters
obtained by extracting image parameters
using a pre-trained neural network and
the Mean Shift algorithm.
Author: Štěpán Miklánek
104
ral networks pre-trained on a freely available image dataset (ImageNet)
containing images aggregated from photo-sharing sites (Flickr.com, etc.),
thus significantly weakening the software’s credibility as a non-human
entity. The software functioned as a tool to search for relationships be-
tween the images from the coronavirus pandemic collected in our Black
box and the images from ImageNet. Since the Black Box dataset is tiny
compared to ImageNet, the neural networks placed more emphasis on
those visual features identical to ImageNet content when sorting our da-
taset. As a result, the arrangement of the photographs remained largely
the same, allowing people, for instance, to mistake selfie images with and
without a mask. When sorting the images, this made the layout of them
a more critical factor than the mask itself, the quarantine symbol. In this
context, the pandemic experience appeared as a banal episode floating
on the surface of the ocean of the visual overproduction of contemporary
visual culture.
105