Due to the rapid development and growth of Artificial Intelligence in recent times,
it is also being used more and more frequently in industrial environments, however
there are many problems besides its advantages. One of them is the difficult process
of Dataset construction which often involves manual sample annotation. This task is
often performed by human labor at larger companies with relative ease, smaller projects
or developments however usually lack the necessary resources to do so, thus making
it seem advantageous to use Artificial Intelligence, more precisely Deep Learning,
for this problem too. Another issue is the asymmetry of these Datasets, as with production
lines, the chance of failure is kept very low, resulting in only a few error samples.
To overcome these issues and further democratize AI, in this paper, we present 2 latent
space-based sampling methods, which based on the acquired results can achieve reasonably
good classification performance with only annotating a few 100s of samples. Moreover,
to aid in the understanding of the results, the presented techniques are compared
to a naive method representing the baseline approach.
