Re:What a machine learning tool that turns Obama white can t...
by clircle on Saturday June 27, 2020 @07:02PM
(#23665393)
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.
Re:What a machine learning tool that turns Obama white can t...
by currymj on Saturday June 27, 2020 @07:37PM
(#23665589)
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.
Re:What a machine learning tool that turns Obama white can t...
by zeroimpl on Saturday June 27, 2020 @09:22PM
(#23666071)
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.
Re:What a machine learning tool that turns Obama white can t...
by clircle on Saturday June 27, 2020 @07:50PM
(#23665649)
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.
Re:What a machine learning tool that turns Obama white can t...
by mc32 on Saturday June 27, 2020 @09:41PM
(#23666166)
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.
With bias, it's helpful to define what property is desired, and then measure the discrepancy from that.
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.
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.
I'm not sure what you mean when you said the model defaults to white features.
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.