After reading the passage "Excavating AI," I immediately connected it to a book I read called Artificial Unintelligence by Meredith Broussard. When training an image-related model, we need to feed it with labeled images and classify them under certain categories. However, labeling images is much harder than we tend to think. As the passage "Excavating AI" states, "Images are remarkably slippery things, laden with multiple potential meanings, irresolvable questions, and contradictions" (Crawford, Paglen). AI is not as “intelligent” as its name suggests. People use data to train models, and data itself is socially constructed. All the data we see is created and generated by people. Every interaction we have with a computer is written by humans who have feelings, biases, and backgrounds. While data is a great and effective tool for analysis, the data-driven approach overlooks many factors that humans consider highly important. A computer can’t identify the extra information that matters, but we can. Humans have judgment, but machines do not. Therefore, decisions made by a model are entirely dependent on the dataset it was trained on. If the dataset is biased, the model’s decisions will be biased as well.
When I looked at the sample images for ImageNet, many of them had more context than their labels suggested. For example, one image was labeled as "bow_tie," but it was clearly about more than just the bow tie—it had much more context than the label implied.
For the first try, I used a picture of my dog Millie, a Maltese. The labels given by the model were actually correct; however, the confidence score was only 0.45, which is not very high. The model wasn’t quite certain about the label it provided. Then, I used a picture of a Maltese dog I found online, and the confidence score increased to 0.7.
However, when I used another picture of Millie for the model to classify, it labeled her as a "teddy" with a low confidence score of 0.25. I believe the model failed to recognize her because the image didn’t show a clear view of Millie's face, and my dad was holding her in the background, so she wasn’t standing or lying down. Additionally, Millie's short haircut makes her resemble a teddy bear dog more.
I also tried the video classifier, and it kept saying, "I see hair spray," while I was just putting my palm in front of the camera. There was definitely no hair spray in the scene, which I found both interesting and a little creepy.
