Reading: “Excavating AI: The Politics of Images in Machine Learning Training Sets”

• Argue that image data sets used to train AI (like object or face recognition) is not neutral and it is deeply social and political
• “archaeology of data sets”
• Images are subjective
  • Everyone has different interpretations of different images (so they depend on context and culture)
  • automated image interpretation simplifies those nuances
• The three layers of dataset construction
  • Taxonomy: what categories exist
  • classes / individual categories: what is presumed in defining these as distinct
  • labeled image themselves: sources of images and how they are selected
• The datasets (and therefore the AI systems trained on them too) magnify social norms and biases
• normative patterns of life are assumed (what is a person, what identities legible, what emotions are valid)

Reflection

Before reading this article and even coming to this class, I never really gave much thought into the idea of datasets being used to train AI models. I used to think datasets were quite neutral in the way that we just give a model a set type of data and it will train off of it. However, I realized that, just like humans, machines are biased because of the information us humans give them. So can anything be truly unbiased? The ImageNet (as seen in the reading) contains offensive and sexist terms in its dataset, and it sort of just exemplifies harmful assumptions on a quiet way which can be detrimental to our society in the long term. I also like how the writer connected ML practices to older forms of race science, because this shows that AI isn’t just making mistakes, it is repeating old ones., just like how some datasets today still try to assign labels to race, gender and emotion as if those are objective truths. At the end of the day, AI is definitely entangled with power and discrimination, and I think it would very challenging to disassociate them from each other.

Training Model (Assignment part 2)

I trained my own Teachable Machine sound model and used it in a p5.js + ml5.js sketch to move a little “blob” with my voice. Saying up / down / left / right / stop / go directly controls the blob’s velocity on the canvas.

I used Teachable Machine → Audio Project and created these classes:

• Up, Down, Left, Right, Go, Stop, Background Noise
• I recorded short clips per label (varied pitch/speed a bit).
• For Background Noise, I captured room ambience