Page 105 - The Black Box Book

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

104 105