3 ms·

What a machine learning tool that turns Obama white can tell us about AI bias

12 points by edward 6y ago

clircle 6y ago
What does it mean that an AI is unbiased? Would it mean that the AI classifies race correctly on average? Or something else -- e.g. zero misclassification error for a specific race? With bias, it's helpful to define what property is desired, and then measure the discrepancy from that.
- currymj 6y ago
  for classifiers, often people want error rates not to differ too much on different groups, but this is fraught with difficulties. it seems to me that it is really situation dependent, complex, and you hit tough philosophical issues about the nature of fairness and justice pretty quickly. on the other hand, with this model, i’m willing to say “i know it when i see it”. if the model did not default to white facial features, it would be less biased.
  - clircle 6y ago
    Another issue: in some machine learning models, like image classifiers, race is not an a priori covariate or group in the data, but is only scrutinized a posteriori. I'm not sure what you mean when you said the model defaults to white features.
  - zeroimpl 6y ago
    It seems like biases related to error rates in image classification systems would be one of the easier problems to address - you basically just need to load in more data until the minimum accuracy level for all peoples is good enough. The harder biases to address are going to be ones where the AI reinforces current undesirable patterns. Eg statistically certain minority groups are more likely to commit petty crimes. If you replaced police with AI robots, these robots would then automatically label people from those minorities as more suspicious. That sucks, and likely has much more complex solutions.
mc32 6y ago
I think Lucy Liu is a much better example of algorithmic bias. Obama is the child of a white parent and a black parent, so will have characteristics of both parents; so the algorithm should have either outcome as a possibility. The default categorization of that mix by people depends on the culture categorizing the person.

12 points by edward 6y ago

What does it mean that an AI is unbiased? Would it mean that the AI classifies race correctly on average? Or something else -- e.g. zero misclassification error for a specific race? With bias, it's helpful to define what property is desired, and then measure the discrepancy from that.

currymj 6y ago
for classifiers, often people want error rates not to differ too much on different groups, but this is fraught with difficulties. it seems to me that it is really situation dependent, complex, and you hit tough philosophical issues about the nature of fairness and justice pretty quickly. on the other hand, with this model, i’m willing to say “i know it when i see it”. if the model did not default to white facial features, it would be less biased.
- clircle 6y ago
  Another issue: in some machine learning models, like image classifiers, race is not an a priori covariate or group in the data, but is only scrutinized a posteriori. I'm not sure what you mean when you said the model defaults to white features.
- zeroimpl 6y ago
  It seems like biases related to error rates in image classification systems would be one of the easier problems to address - you basically just need to load in more data until the minimum accuracy level for all peoples is good enough. The harder biases to address are going to be ones where the AI reinforces current undesirable patterns. Eg statistically certain minority groups are more likely to commit petty crimes. If you replaced police with AI robots, these robots would then automatically label people from those minorities as more suspicious. That sucks, and likely has much more complex solutions.

mc32 6y ago

I think Lucy Liu is a much better example of algorithmic bias. Obama is the child of a white parent and a black parent, so will have characteristics of both parents; so the algorithm should have either outcome as a possibility. The default categorization of that mix by people depends on the culture categorizing the person.