Page 103 - The Black Box Book

Embedding) algorithm was chosen, allowing the reduction of numerical
vectors to points in a two- or three-dimensional space. (Pouyet, 2018)
This way, each of the input images was given a precise location on a map
of the similarity-difference relations between the images in the dataset
in the form of two coordinates. This metadata subsequently determined
the location of the image on the similarity map (Figure 6), which was
the product of this experiment.
However, it turned out that the colour parameter as the domi-
nant criterion for dataset classification is not suitable for our purposes.
Artists often submitted images that were similar in colour, naturally, as
they were taken in the same or similar environment or additionally modified
with the same filter. As an outcome, the first experiment’s results showed
high accuracy in classifying images according to the creators without
equipping the neural networks with a model of the authors’ handwrit-
ing or characteristic motifs. On the one hand, in this case, the artificial
neural network proved to achieve high accuracy in sorting the dataset
according to authors based on minimal input information (in our case,
the dataset was sorted according to a single criterion – the colour spec-
trum used). However, for our purposes, this proved to be unsatisfactory,
Figure 6: Image output of attempt 1 as we wanted the non-human (alien) curator to look for recurring motifs across
– Visualization of image similarity using t-SNE algorithm. the dataset.
Author: Štěpán Miklánek
102 103