There are a few things that can be said about the ethical implications of training deep learning models. There is potential harm whenever you are handling data that represents human perceptions. But also, data about humans and human interaction has to be rigorously protected and examined carefully before creating a model that will generalize based on such data. Such thoughts are organized in the following sections.
Reporting using the appropriate performance measures
Avoid faking good performance by picking the one performance metric that makes your model look good. It is not uncommon to read articles and reports of multi-class classification models that are trained over clear, class-imbalanced datasets but report the standard accuracy. Most likely, these models will report a high standard of accuracy since the models will be biased toward the over-sampled class and against the under-sampled groups. So, these types of models must...