3 ms·

> If you use basic statistical methods and realize that murders are more likely to occur in this location between these hours and adjust your police force accor

by Closi 5y ago

> If you use basic statistical methods and realize that murders are more likely to occur in this location between these hours and adjust your police force accordingly, is this bad? It depends - These models are often non-transparent in terms of how they calculate, so what if the AI/model decides/calculate that a higher percentage of minorities in an area is a high predictor of crime, thus indirectly creates a policy that, for instance, areas with high levels of ethnic minorities require higher levels of policing or harsher restrictions (i.e. curfews, more CCTV e.t.c.). If it's just a basic 'well murders happen between these hours on this street so we will put a police officer there' then of course that should happen, but that's not really policing based on AI/ML, it's just based on basic statistics.

_ke2e 5y ago
Where do you draw the line between statistics and ML?
- brightly-salty 5y ago
  When you can no longer transparently explain the reasoning behind a decision made based on the statistics, and instead have to just shrug and say “because the model told us so”. Accountability is very important in this issue.
  - bko 5y ago
    Apart from neural networks most models are somewhat transparent in terms of reasoning. Neural networks would be the wrong model to use for this anyway. For instance, decision trees can tell us if the population density between this range and the percentage of single family residence is between this range, the probability of a package being stolen is X. Or regression could tell us that the square of the population density is positively correlated to this crime .
    - Closi 5y ago
      This law isn’t talking about banning decision trees or banning regression analysis though, it’s talking about ML being used in specific listed applications.
    - brightly-salty 5y ago
      Exactly. We don’t disagree, I don’t think. Decision trees and regression analysis should count as statistics for the purpose of this issue, as they are transparent and accountable. Neural networks are another beast entirely, in my opinion.
- Closi 5y ago
  I think they are two different things - statistical modelling to me is about understanding relationships between data (i.e. we are highly confident that X and Y are correlated, and can infer a reason and thus put a policy around that based on open reasoning), and ML is about making predictions in new situations based on historical training data from different situations (but we can't necessarily infer the why).
bko 5y ago
> It depends - These models are often non-transparent in terms of how they calculate, so what if the AI/model decides/calculate that a higher percentage of minorities in an area is a high predictor of crime, thus indirectly creates a policy that, for instance, areas with high levels of ethnic minorities require higher levels of policing or harsher restrictions (i.e. curfews, more CCTV e.t.c.). It's pretty simple to make the models transparent. Just select the inputs you want the model to have access to. You shouldn't be feeding your model the percentage of minorities in an area. What's not transparent is how humans make decisions. You can't tell a human to consider only these characteristics but not these others ones. > If it's just a basic 'well murders happen between these hours on this street so we will put a police officer there' then of course that should happen, but that's not really policing based on AI/ML, it's just based on basic statistics. I don't see how you make a meaningful separation between the two. Either you can use predictive models or not. It's hard to carve out "use predictive models but not too predictive". The more likely scenario is just not using any model and distributing police resources uniformly regardless of need.
- Closi 5y ago
  > It's pretty simple to make the models transparent. Just select the inputs you want the model to have access to. You shouldn't be feeding your model the percentage of minorities in an area. That doesn't make the model transparent, it just makes the input and output transparent, but it's still a black box. It's nice to say that everyone will ensure that the model isn't fed with information which will allow it to make decisions or be biased based on sex / gender / race / income but in reality accidental racial bias has already shown up multiple times in applications where it should be easier to detect than an invisible 'crime prediction' algorithm that the general public won't have access to but public policy will be implemented on the basis of. > I don't see how you make a meaningful separation between the two. Either you can use predictive models or not. It's hard to carve out "use predictive models but not too predictive". The more likely scenario is just not using any model and distributing police resources uniformly regardless of need. I make a meaningful separation because one has clear logic which can be viewed to be unbiased, and the other has a black box calculating it.
  - 5y ago
throwawaysea 5y ago
I am still not clear on why that's a problem. If the goal is to reduce or deter crime, and there's an effective way to do that, why does it matter what correlation it has to outside demographic factors? I would argue that not containing crime in a known high-crime area just because of correlation to 'high levels of ethnic minorities' is itself racist, since that is itself an act of making race a factor in decision making. It also doesn't fit what minorities actually want. As an example, an overwhelming majority of Black Americans want policing to stay the same or increase (https://news.gallup.com/poll/316571/black-americans-police-retain-local-presence.aspx https://news.gallup.com/poll/316571/black-americans-police-r...), rather than decrease, despite what (predominantly-white) activists have been pushing for in the last year.

by Closi 5y ago

_ke2e 5y ago

Where do you draw the line between statistics and ML?

brightly-salty 5y ago
When you can no longer transparently explain the reasoning behind a decision made based on the statistics, and instead have to just shrug and say “because the model told us so”. Accountability is very important in this issue.
- bko 5y ago
  Apart from neural networks most models are somewhat transparent in terms of reasoning. Neural networks would be the wrong model to use for this anyway. For instance, decision trees can tell us if the population density between this range and the percentage of single family residence is between this range, the probability of a package being stolen is X. Or regression could tell us that the square of the population density is positively correlated to this crime .
  - Closi 5y ago
    This law isn’t talking about banning decision trees or banning regression analysis though, it’s talking about ML being used in specific listed applications.
  - brightly-salty 5y ago
    Exactly. We don’t disagree, I don’t think. Decision trees and regression analysis should count as statistics for the purpose of this issue, as they are transparent and accountable. Neural networks are another beast entirely, in my opinion.
Closi 5y ago
I think they are two different things - statistical modelling to me is about understanding relationships between data (i.e. we are highly confident that X and Y are correlated, and can infer a reason and thus put a policy around that based on open reasoning), and ML is about making predictions in new situations based on historical training data from different situations (but we can't necessarily infer the why).

bko 5y ago

> It depends - These models are often non-transparent in terms of how they calculate, so what if the AI/model decides/calculate that a higher percentage of minorities in an area is a high predictor of crime, thus indirectly creates a policy that, for instance, areas with high levels of ethnic minorities require higher levels of policing or harsher restrictions (i.e. curfews, more CCTV e.t.c.). It's pretty simple to make the models transparent. Just select the inputs you want the model to have access to. You shouldn't be feeding your model the percentage of minorities in an area. What's not transparent is how humans make decisions. You can't tell a human to consider only these characteristics but not these others ones. > If it's just a basic 'well murders happen between these hours on this street so we will put a police officer there' then of course that should happen, but that's not really policing based on AI/ML, it's just based on basic statistics. I don't see how you make a meaningful separation between the two. Either you can use predictive models or not. It's hard to carve out "use predictive models but not too predictive". The more likely scenario is just not using any model and distributing police resources uniformly regardless of need.

Closi 5y ago
> It's pretty simple to make the models transparent. Just select the inputs you want the model to have access to. You shouldn't be feeding your model the percentage of minorities in an area. That doesn't make the model transparent, it just makes the input and output transparent, but it's still a black box. It's nice to say that everyone will ensure that the model isn't fed with information which will allow it to make decisions or be biased based on sex / gender / race / income but in reality accidental racial bias has already shown up multiple times in applications where it should be easier to detect than an invisible 'crime prediction' algorithm that the general public won't have access to but public policy will be implemented on the basis of. > I don't see how you make a meaningful separation between the two. Either you can use predictive models or not. It's hard to carve out "use predictive models but not too predictive". The more likely scenario is just not using any model and distributing police resources uniformly regardless of need. I make a meaningful separation because one has clear logic which can be viewed to be unbiased, and the other has a black box calculating it.
- 5y ago

throwawaysea 5y ago

I am still not clear on why that's a problem. If the goal is to reduce or deter crime, and there's an effective way to do that, why does it matter what correlation it has to outside demographic factors? I would argue that not containing crime in a known high-crime area just because of correlation to 'high levels of ethnic minorities' is itself racist, since that is itself an act of making race a factor in decision making. It also doesn't fit what minorities actually want. As an example, an overwhelming majority of Black Americans want policing to stay the same or increase (https://news.gallup.com/poll/316571/black-americans-police-retain-local-presence.aspx https://news.gallup.com/poll/316571/black-americans-police-r...), rather than decrease, despite what (predominantly-white) activists have been pushing for in the last year.