The rapid integration of machine learning into our society has emerged the
concerns about its potential discrimination in terms of gender and race caused
by the unconscious bias introduced in it, which has led to the necessity to
incorporate new sustainable practices to reduce its societal implications.
Whilst the issue of bias has been proposed as an essential subject to be
included in the new AI curricula for schools, research has shown that it is a
difficult topic to grasp by children. This study aims to develop an
educational platform tailored to raise the awareness of the societal impacts
of gender bias in supervised learning. Building on existing work on
interpretable machine learning, this thesis is focused on the specific
research question assessing whether using an interpretable model in an
educational platform has an effect in preadolescents from 10 to 13 years old
understanding the concept of bias in these systems and its ethical
implications. In this context, an interpretable model is conceived as that
able to provide an explanation to its own predictions.
A study was carried out
in a school in Stockholm employing an online platform developed to include an
interpretable model using the Grad-CAM explanation technique to run the
experiment. The students were divided into two conditions, differentiated by
the use of the explainable model or not, and their understanding of bias was
evaluated on an interactive activity where they had to select the most
relevant parts of an image used by the model after observing its prediction,
along with other multiple choice questions before and after the main activity.
Analysis of the answers demonstrated a significant effect when using the
interpretable model, improving the ability of preadolescents to recognize the
impact of bias in terms of gender discrimination, and identify training data
as an essential part in this issue. These results indicate that our proposal
of using explainability techniques has an impact on the understanding of bias
in machine learning and its ethical implications, providing evidence of its
effectiveness to be used in future educational platforms, and potential to be
introduced into the new AI curricula guidelines, specifically aimed to raise
the awareness of the societal implications of bias in machine learning.