Reflection on Excavating AI: The Politics of Images in Machine Learning Training Sets by Kate Crawford and Trevor Paglen

<aside> <img src="/icons/question-mark_blue.svg" alt="/icons/question-mark_blue.svg" width="40px" />

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

1. Anyone has the power to label images, it just depends who uses the dataset.
2. Models may be trained on data that is biased and underestimates the contextual weight of a noun. </aside>

Notes

Building a computer vision system:

1. Collect images
2. Label images
3. Train neural network

“Entire subfields of philosophy, art history, and media theory are dedicated to teasing out all the nuances of the unstable relationship between images and meanings.”

• ImageNet by Fei Fei Li (Stanford)
  • Based on the semantic structure of WordNet (database of word classifications - Princeton)
• Amazon Mechanical Turk

The problem with ImageNet

• A noun usually falls under a spectrum between…
  • concrete ↔ abstract
  • descriptive ↔ judgmental
• However, ImageNet disregards space for context (societal, judgmental, and cultural biases) and concepts become things of concrete truths
  • Abstractions turn into hard categories
  • Dangerous when applied to humans
    • Imagine applying derogatory terms to a subset of people
• Does not touch the subject of Physiognomy ”judging someone’s personality, intelligence, or morals based on facial features”
• Disregards the political history of diversity