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?
The power to assign labels to images often rely on individuals designated by dataset creators. Relying on human classification often reflects the biases, perception, and cultural contexts of society — or the priorities of the company creating the dataset. Labeling is not objective so the act of categorizing and defining an image involves subjective choices which may enforce rigid contemporary biases. Magritte’s painting mirrors this concept, pointing out how representation conflicts with reality and cannot always be trusted in the way we understand the world.
If the labels in the training data carries bias the resulting models will perpetuate these biases and force them upon the users which only worsens from there.
I started by trying out with a couple of different objects. I had a bunch of image samples for each object and they were all pretty different objects, so the model seemed to detect the classes pretty well.
I wanted to try to get the model to recognize colors. I trained the model on a couple different objects of each color and put it color classes.
The goal was that if I showed the model something which I didn’t train it on, it would be able to classify it based on color. I noticed this worked the best for red, likely because of how bright and similar the red is. I trained it on the red box, and it worked on the tylenol cap as well. However this wasn’t as effective for green and blue objects
Then I decided to turn this model into a character color selector, so I just added a character with p5.js and used the class as a string for the fill function.