Page 107 - The Black Box Book

Figure 8: Structure of
ATTEMPT 3: ALIEN CURATOR the neural network – autoencoder.
In the third experiment, a neural network that was not pre-trained on Author: Štěpán Miklánek
any visual dataset was used. This type of neural network is referred to as
an autoencoder (Hartman & Mestha, 2017). This neural network architec-
ture is used for dimensionality reduction. The task of the autoencoder is
to approximate the function between the input and output datasets. This
architecture is divided into two sides i.e., encoder and decoder. The first
block (encoder) reduces the image input into a simplified representation
so that the original image can be reconstructed from this representation.
The second block (decoder) is used to reconstruct the original image
back from the simplified representation of the original digital image.
In this process, the neural network reduces the visual data of the original In this experiment, we were able to come closest to what we think
image into a simplified representation that captures the characteristic of as an alien curator, as we minimized the contact of the developed
parameters of the image. (Figure 8) The output of the encoder was then neural network with human-generated inputs or decisions, whether in
used as input to the clustering algorithm. (Hinton, 2006) The decoder the form of a metadata-laden dataset created by humans, such as
was only used in the training process to verify the reduced visual data. the aforementioned ImageNet, or by programmers of a predefined num-
As in the previous case, the Mean Shift algorithm was used at this stage. ber of clusters or other parameters to guide the software’s performance.
The result was the classification of the image dataset into thematic groups As the dataset expands and changes in the future, the neural network
using machine learning based on the recognized similarities derived solely will also evolve, which may cause, for example, changes in the number of
from the machine analysis of the respective dataset. clusters (thematic groups) of recognized objects.
The clusters of images the neural network has evaluated as similar
to each other contain images that, from our point of view, are quite het-
erogeneous, often showing no common denominator, a shared element,
whether in the form of an object or composition. We can only guess what
parameters were taken into account when classifying them. We are thus
confronted with a non-human entity and its calculating logic that performs
the visual analysis of the image traces of the coronavirus pandemic.
106 107