Colin’s Reviews > The Alignment Problem: Machine Learning and Human Values > Status Update

DL systems are brittle - they MUST classify even static as one thing or another

We don't want algorithmic recommenders naively reinforcing our worst impulses. We are not perfect.

RL - telling AI what we want (coding reward function)
IRL - showing AI what we want (having them infer from model behavior)
Synthesis - knowing what we want when we see it and communicating that to the AI

Dopamine - currency of reinforcement learning in the brain, telling you that your expectations are in error and you should learn from it. The reason why cocaine is addictive - you feel a rush from thinking it's gonna get better, but that's just the artificial dopamine speaking.

Predictive policing - what if model predictions determine policing rates? A selection bias arising from poor quality data might tag a place as ok when it is not, resulting in reallocation of police resources away, exacerbating crime. Initiating a feedback loop. Selection bias meets confirmation bias.

It is impossible to guarantee equal fairness across all domains of errors in a classification algorithm if the base rates between categories are not themselves equal. If you optimize for false negatives you will lose on false positives. That means all algorithms will be open to critique and you need to optimize for what is more fit for purpose

Being bias-blind doesn’t work for fairness - not only are correlated variables still encoded regardless, removing these variables from analysis makes it harder to correct for reality-bias. The question is not what data is in the set but how it is used.

It is not the algorithm at fault but the representativeness of the training dataset