Reflect on the relationship between labels and images in a machine learning image classification dataset.

Who has the power to label images and how do those labels and machine learning models trained on them impact society?

I had several classes and reading groups talking about AI and its bias before, so I understood the severity it is to future human civilization development. However, even with a bit background knowledge, I was impressed by the examples and the timeline of the development in image classification.

The phrase “machine learning”, or to be more specific the word “learning”, will confuse many who don’t have any knowledge in computer science that the learning for computers or models is the similar to us human beings. However, just as the authors mentioned at the very beginning of the article, “teaching” or training computers to identify and classify images is very hard as computers only have sensors and we need to first teach term to “see” things.

I agree that labeling is an effective way to do image classification as we humans nowadays labeling a lot of things based on our opinions. And labels are useful in images such as, as authors mentioned, “apple” and “apple butter”. It is impossible to label “apple” as a, for example, a dog, because such nouns don’t contain any subjective idea. So under this situation, starting with labeling is a good choice as researchers tend to begin with these simpler nouns.

However, when the training goes deeper where simple nouns are all well-trained, it is inevitable that they will begin to include more subjective nouns, and this is when things become complicated and more politics get involved.

Nowadays, researchers, dataset curators, and annotators have the right to label images. And as authors mentioned, such classification inevitably contains subjectivity as labels like “happy”, or “healthy man” is different to each person.

The models trained on them can have impact on the good side, with more efficient decision-making, medical assistance, accessibility support, and automation in various industries. However, they can also reinforce stereotypes, amplify biases, and even cause harm when used in sensitive areas such as hiring, policing, or healthcare. Since models inherit the worldview of the labelers, the power of deciding what label to apply translates into real influence over how people are represented, treated, and understood in society. This is why transparency, diversity of perspectives, and careful governance are essential in the process of dataset labeling and model training.