Abstract
Convolutional neural networks are one of the models which are introduced at the beginning when starting to learn about machine learning. While visualizations exist for
teaching about the concepts and for improving the model after it has run, there are no
visualizations while learning to code. Therefore, the question arises whether having a
visualization during the building of a model is useful. To measure usefulness, the two
research questions are “How does visualizing a self-built CNN contribute to students’
understanding of CNN concepts after exploring a predefined model?” and “How do stu
dents perceive the contribution of constructing and visualizing a CNN model to their
understanding of CNNs and to their confidence in coding them?”. CodeVisCNN, a web
application that generates layer animations of the given model code, was developed for
the study. Seventeen participants did a pre-, mid-, and post-test about convolution layer
knowledge. Between the pre- and mid-test, they interacted with a CodeVisCNN version
that only showed a predefined model, while the full version was available between the
mid- and post-test. In the end, every participant completed a questionnaire. The results
show that the change from the predefined to the student’s own model contributes to
their understanding because from the mid- to post-test the number of correct answers
and their confidence in their answer increased statistically significantly while the com
pletion time decreased statistically significantly. While the participants’ opinion about
the biggest contributor to their understanding is nearly equally distributed between
both, the first and second version, the majority declared the second visualizer as the
biggest contributor to their confidence in writing.
| Date of Award | 2026 |
|---|---|
| Original language | English (American) |
| Awarding Institution |
|
| Supervisor | David Christian Schedl (Supervisor) |
Studyprogram
- Interactive Media
Cite this
- Standard