Objectives

• Define Artificial Intelligence (AI), Machine Learning (ML), and Deep Learning (DL).
• Understand what ml5.js is and how it fits into the TensorFlow and open source machine learning library ecosystem.
• Learn how to create an image classifier with ml5.js and MobileNet.
• Understand how the MobileNet model was trained, specifically the origins and collection methodology for the training.
• Distinguish between a “feature vector” (aka logits) and the last layer (aka softmax probabilities) of a classification network.
• Understand the process of “transfer learning”.

Learning Materials

• https://github.com/shiffman/ML-for-Creative-Coding/tree/main/01-introduction
• Excavating AI: The Politics of Images in Machine Learning Training Sets
• Humans of AI

Work Desk

Reflection on Readings

• political landscape is too volatile
  • rate of change > time it takes to collect annotated(labeled) data (e.g. I was first introduced to gender as a spectrum 7 years ago and was utterly shocked)
  • the dataset is a snapshot of mass opinion/publicly available data at that moment - it’s not surprising that it looks shocking now but the labeling started 10 years ago
  • also I believe that liberal opinion is overrepresented in media but the people who are involved in this task may pose a different political stance. I think the imbalance between the tone between media and the actual technology rises from this
• the labeled dataset is subject to sampling error of:
  • those who define the categories
  • those who label data
    • akin to who participates in clinical drug trials. a certain socioeconomic class, education level
• also it’s a very quick job per image(=cheap data)
  • unlikely that many cross-check labels for a single image (likely that samples were cross-checked if it ever happened) and it’s a task that takes less time making it even more prone to those tasked revealing implicit associations
• solutions?

  • educate the responsibility of all parties involved in the generation / collection / definition / classification of data

    • I don’t think those involved in the people category of ImageNet fully grasped what implications their tasks had from the onset
    • in this context I think the Humans of AI is a good piece of writing that communicates the weight of their responsibilities emphatically

  • attach more weight to such decisions (e.g. multiple ppl labeling data and averaging, etc)

    ⇒ makes data more expensive - which goes against the current hypercompetitive, hyper growth market trends