Lecture 38 - Bias in ML

Announcements:

• Reminder: data ethics 3 discussion tomorrow (Wednesday), with in-class graded writing prompt.
• Lab 9 final submission deadline 4(?)pm today; Canvas submission deadline 10pm tonight.
• Please fill out course evals!
• Friday's class: optional, AMA in AW 303
• Details for FP final submission out soon

Goals:

• Gain an appreciation for how bias can manifest in machine learning systems

Now that you know the overall shape of how to train and use machine learning models, it's worth asking:

What could possibly go wrong?

• out of distribution
• self-driving cars hit pedestrians
• automated stock trading
• label imbalance -> accuracy imbalance
• ML just doesn't work
• train on unrealistic / inaccurate / biased data

• Criminal justice system (Data Ethics 3)
• Face recognition systems: http://gendershades.org/overview.html

The state of the art, especially in vision and language

• More data --> better performance.

  • Example: image recognition started working in large part because of ImageNet, a dataset of around 14 million labeled images. This dataset is now considered 'small'.
  • Example: GPT-3, state-of-the-art language model trained on a corpus of 500 billion words from all across the internet

• One problem:

  • ImageNet images were labeled by humans
  • The internet was written by humans

Example: how-to-make-a-racist-ai-without-really-trying.ipynb

Further reading:

ConceptNet Numberbatch 17.04: better, less-stereotyped word vectors

Semantics derived automatically from language corpora contain human-like biases

Image Representations Learned With Unsupervised Pre-Training Contain Human-like Biases

Towards Fairer Datasets: Filtering and Balancing the Distribution of the People Subtree in the ImageNet Hierarchy

racist data destruction? a Boston housing dataset controversy