Fixing bias in AI
Nowadays several techniques exist for countering bias in AI applications. In practice, that means: countering bias in applications using forms of machine learning.
TODO: link notes on techniques for fairness and countering bias. Such as SHAP
But ...
I argue that these tools should not be treated like plug-and-play tools to "fix" bias in machine learning pipelines. Their correct use should first of all be embedded in a broader analysis of where the bias comes from and what's the best way to tackle it.
I can image there can be different locations in a machine learning pipeline where you may intervene. I came up with these ideas before a thorough study of the available techniques, but these are my initial thoughts:
- You may intervene already in the data collection step.
- You may intervene before training by introducing additional pre-processing..
- You may develop techniques to identify bias in trained network and then compensate for this bias during prediction time.
But bias in AI is a socio-technical problem. As a result, it is not sufficient to only apply technical fixes that fix an algorithm. You also need to check the context of usage. Tackling bias could then involve several things:
- To not apply an algorithm in certain usage contexts where it is likely to disproportionally affect a particular group of people
- The decision to not use or develop a technique should be on the table