4 ms·

Without links ready, and not being the OP, I have seen many articles saying that "this ML recognition/analysis system has racial bias because the targets it fin

Without links ready, and not being the OP, I have seen many articles saying that "this ML recognition/analysis system has racial bias because the targets it finds are more {insert trait} than average." It struck me how such an unscientific thing made it to so many articles. It's a symptom of a cause that a specific group has greater representation in something, e.g. crime. You can't call factual observations, racist. You should rather find the root cause and solve it.

vmception 6y ago
> You can't call factual observations, racist. You should rather find the root cause and solve it. People already know the root cause, and it is overpolicing and discretionary enforcement of crimes like drug possession. The broken-windows policy has been disproven, despite disproportionately impacting "the specific groups" (black Americans) that now pollute the dataset that ML uses. And even in the concept or drug possession and drug consumption, all groups have been shown to use them in the same distribution. For example. These kinds of things start a cycle that means the second and third minor infractions cause greater consequences in court, which further reduce opportunities that lead to the dangerous crimes being committed. So we already know the dataset is polluted. Pointing to the top of the iceberg and saying "well they commit the crimes no need to spend any energy on this mystery why don't we all just admit they're the problem" is a complete deflection promulgated intentionally and you should really check your peer group and media sources if this is the extent of the comfortable worldview you (or anyone passing by) have, the problem and solution is already known. The solutions are being implemented in a patchwork and slowly, which is not incorporated in datasets that ML use, largely due to apathy and lack of awareness of engineers and the management of the tech companies involved, and lack of representation of the affected groups in engineering and management of tech companies.
- core-questions 6y ago
  > People already know the root cause, and it is overpolicing and discretionary enforcement of crimes like drug possession. No, you've only pushed the root cause back a step by doing this. It's entirely possible to understand that the way we're doing policing is wrong without resorting to an artificial, borderline-creationist mindset on aggregate group behaviour. Religious thinking is not going to help anything here. > So we already know the dataset is polluted. The dataset reflects something approaching reality. If anything, it reflects a version of reality that is already attempting to artificially compensate for group differences in order to quell conflict. If you want to change reality, if you want to see less crime, if you want to see truly fair policing, then admitting to reality is an important first step. You're basically arguing for "juking the stats" in order to find fairness, when in reality such actions won't stop people from getting robbed, murdered, or having lives that offer so few opportunities for advancement that they end up turning to hard drugs to cope. ML is not to solve this problem either way, but it can actively prevent the problem from being solved if it becomes yet another mechanism to paper over the actual situation and instead point the finger for responsibility away from where it belongs: with individuals and their choices.
spoonjim 6y ago
Well the “factual” observation comes from training data, which unless created in an unbiased manner creates a biased dataset and biased inferences. Unless you believe that American law enforcement is an unbiased process, in which case I don’t think I can help you.

vmception 6y ago

> You can't call factual observations, racist. You should rather find the root cause and solve it. People already know the root cause, and it is overpolicing and discretionary enforcement of crimes like drug possession. The broken-windows policy has been disproven, despite disproportionately impacting "the specific groups" (black Americans) that now pollute the dataset that ML uses. And even in the concept or drug possession and drug consumption, all groups have been shown to use them in the same distribution. For example. These kinds of things start a cycle that means the second and third minor infractions cause greater consequences in court, which further reduce opportunities that lead to the dangerous crimes being committed. So we already know the dataset is polluted. Pointing to the top of the iceberg and saying "well they commit the crimes no need to spend any energy on this mystery why don't we all just admit they're the problem" is a complete deflection promulgated intentionally and you should really check your peer group and media sources if this is the extent of the comfortable worldview you (or anyone passing by) have, the problem and solution is already known. The solutions are being implemented in a patchwork and slowly, which is not incorporated in datasets that ML use, largely due to apathy and lack of awareness of engineers and the management of the tech companies involved, and lack of representation of the affected groups in engineering and management of tech companies.

core-questions 6y ago
> People already know the root cause, and it is overpolicing and discretionary enforcement of crimes like drug possession. No, you've only pushed the root cause back a step by doing this. It's entirely possible to understand that the way we're doing policing is wrong without resorting to an artificial, borderline-creationist mindset on aggregate group behaviour. Religious thinking is not going to help anything here. > So we already know the dataset is polluted. The dataset reflects something approaching reality. If anything, it reflects a version of reality that is already attempting to artificially compensate for group differences in order to quell conflict. If you want to change reality, if you want to see less crime, if you want to see truly fair policing, then admitting to reality is an important first step. You're basically arguing for "juking the stats" in order to find fairness, when in reality such actions won't stop people from getting robbed, murdered, or having lives that offer so few opportunities for advancement that they end up turning to hard drugs to cope. ML is not to solve this problem either way, but it can actively prevent the problem from being solved if it becomes yet another mechanism to paper over the actual situation and instead point the finger for responsibility away from where it belongs: with individuals and their choices.

spoonjim 6y ago

Well the “factual” observation comes from training data, which unless created in an unbiased manner creates a biased dataset and biased inferences. Unless you believe that American law enforcement is an unbiased process, in which case I don’t think I can help you.