Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Moral Machine

moralmachine.mit.edu

Post reply on HN
Print view View original

31–40 of 48 posts

Re: Moral Machine

#31

Post by domador »

TheSmiddy wrote: post #22
cscurmudgeon wrote: post #17

Earlier quoted context omitted.

Trolley problems are not about trolleys. Don't take them literally (e.g. Schrodinger's cat).

The original trolley problem of "should you let 5 people die or actively kill 1 to save the 5" has genuine philosophical merit, these however are just so contrived they can't possibly offer any meaningful insight while at the same time planting the seed in the general public that this is the type of problem a self driving car should care about which could cause all kinds of wacky regulations to exist.

To me that problem had a simple answer: active harm is worse than negligence. To kill someone is a worse action that to let someone die, regardless of the numbers of people involved.

Re: Moral Machine

#32

Post by sneak »

domador wrote: post #31
TheSmiddy wrote: post #22

Earlier quoted context omitted.

The original trolley problem of "should you let 5 people die or actively kill 1 to save the 5" has genuine philosophical merit, these however are just so contrived they can't possibly offer any meaningful insight while at the same time planting the seed in the general public that this is the type of problem a self driving car should care about which could cause all kinds of wacky regulations to exist.

To me that problem had a simple answer: active harm is worse than negligence. To kill someone is a worse action that to let someone die, regardless of the numbers of people involved.

I personally agree with you, however is is extremely important to realize that most of our society does not agree at all, and that that data is hugely relevant.

Re: Moral Machine

#33

Post by vlovich123 »

whyte_mackay wrote: post #30
TheSmiddy wrote: post #3

I think these trolley problems are a waste of everybody's time. Building redundant reliable braking systems will be orders of magnitudes easier than creating a system to fairly and accurately assess who is the best set of people to kill in a disaster scenario.

I hear this often (that trolley problem is not relevant), but then I discovered that a lot of realistic ML fairness problems can be restated as a trolley problem. You have a classifier for credit assignment (giving a loan, etc.). The classifier is 99% accurate on the entire population. The classifier is 55% accurate on a small minority. You can improve the minority accuracy to 90% at the cost of 0.3% decrease of gene…

Have the classifier incorporate the identifying characteristics of the minority so that it can push past beyond the 99% limit it's at currently.

This is why people are pissed off about the trolly problem. It's all zero-sum hypotheticals when reality is primarily not zero-sum.

Re: Moral Machine

#34

Post by whyte_mackay »

vlovich123 wrote: post #33
whyte_mackay wrote: post #30

Earlier quoted context omitted.

I hear this often (that trolley problem is not relevant), but then I discovered that a lot of realistic ML fairness problems can be restated as a trolley problem. You have a classifier for credit assignment (giving a loan, etc.). The classifier is 99% accurate on the entire population. The classifier is 55% accurate on a small minority. You can improve the minority accuracy to 90% at the cost of 0.3% decrease of gene…

Have the classifier incorporate the identifying characteristics of the minority so that it can push past beyond the 99% limit it's at currently. This is why people are pissed off about the trolly problem. It's all zero-sum hypotheticals when reality is primarily not zero-sum.

Adding fairness to your models usually incurs a cost which can be measured. You have to choose between max profit or equal opportunity, you can't have both (do you run over the investors or over the minorities?).

See here a visualization of trade-off's between global accuracy and fairness:

http://research.google.com/bigpicture/attacking-discriminati...

Re: Moral Machine

#35

Post by polkapolka »

vlovich123 wrote: post #33
whyte_mackay wrote: post #30

Earlier quoted context omitted.

I hear this often (that trolley problem is not relevant), but then I discovered that a lot of realistic ML fairness problems can be restated as a trolley problem. You have a classifier for credit assignment (giving a loan, etc.). The classifier is 99% accurate on the entire population. The classifier is 55% accurate on a small minority. You can improve the minority accuracy to 90% at the cost of 0.3% decrease of gene…

Have the classifier incorporate the identifying characteristics of the minority so that it can push past beyond the 99% limit it's at currently. This is why people are pissed off about the trolly problem. It's all zero-sum hypotheticals when reality is primarily not zero-sum.

The identifying characteristics are signal for the minority, but noisy for genpop.

Deployment/engineering constraints call for a single model.

Realistic scenario.

Re: Moral Machine

#36

Post by chongli »

1999 wrote: post #9

These scenarios are idiotic. If you want to wank off about self driving car ethics, here is a much more realistic scenario: should all self-driving cars report their location to 911 dispatch to allow any vehicle to be re-purposed as an emergency vehicle at any time? That might actually save someone. Also, can anyone identify a useful idea that philosophers have come up with in the last 50 years?

should all self-driving cars report their location to 911 dispatch to allow any vehicle to be re-purposed as an emergency vehicle at any time?

That doesn't make much sense. The primary advantage of emergency vehicles is to transport medics to the site of the emergency so that they can administer first aid and stabilize the person for safe transport to the hospital.

Having just any person off the street pick up a critically injured person is not going to go well. In all likelihood, they'll further injure the person due to their lack of training.

Re: Moral Machine

#37

Post by joe_the_user »

whyte_mackay wrote: post #30
TheSmiddy wrote: post #3

I think these trolley problems are a waste of everybody's time. Building redundant reliable braking systems will be orders of magnitudes easier than creating a system to fairly and accurately assess who is the best set of people to kill in a disaster scenario.

I hear this often (that trolley problem is not relevant), but then I discovered that a lot of realistic ML fairness problems can be restated as a trolley problem. You have a classifier for credit assignment (giving a loan, etc.). The classifier is 99% accurate on the entire population. The classifier is 55% accurate on a small minority. You can improve the minority accuracy to 90% at the cost of 0.3% decrease of gene…

You have a classifier for credit assignment (giving a loan, etc.). The classifier is 99% accurate on the entire population. The classifier is 55% accurate on a small minority.

Uh, concerning this hypothetical.

I'll admit a scenario of this sort sounds appealing at first blush. But "99%" accuracy rate with credit assignment is transparently absurd if considers it for a second. There is a clear, significant limit to the accuracy that can being assigned to anyone's credit, if credit means "actually repaying". The fundamental uncertainty of the economy guarantees this.

The distinction between this ideal (99%-55%) and whatever it might be in reality ( 65%-55%) matters. What's is the system is squeezing a few more percentage points out of data for a large company. And what is the cost of those percentage points?

[EDIT: ACTUALLY - the pernicious scenario is a system that isn't not any MORE accurate for any group than any other BUT which is NEGATIVELY biased against one group and POSITIVELY biased against another group. That situation is EASY to get when one unselectively slurps up any data available. The inaccuracy of predictor is a problem for the company, the biasedness of the predictor is a problem of the individuals discriminated against]

The situation is that a company really can a total better prediction rate for various desired qualities by using completely biased, unfair markers. (White-skin, went to "a good school", from a wealthy background, dresses well, attractive features...). When one allows "black box optimization" to get those features, what one does is allow the use of these considerations, which all otherwise legally off-limits. Legal strictures against discrimination say that objective measures of black people's ability need to be it, not because other measures never matter but because other measures are unfair, other measures don't consider past discrimination.

As a further example, outside of race or gender considerations, some percentage of employees may be forced to care for a sick relative. Maybe that makes them a potentially less effective employee or worse credit risk or whatever. Human evaluators might have values that such questions outside consideration. For an opaque multidimensional analysis, this may a ding - the human user doesn't even know if it's a ding.

Re: Moral Machine

#38

Post by flukus »

baroffoos wrote: post #25
zzzcpan wrote: post #10

Earlier quoted context omitted.

Why would you program a solution that has to kill people? If you are aware of specific situations, shouldn't you program a solution that completely avoids such situations and saves everyone?

Some situations cant be avoided like if something falls off the car in front of you. In a real car you could swerve to avoid it and you will likely kill whoever is on the side of the road you swerve at but you will get away with it because it was a panic response and there was no way for you to properly assess what to do. This changes with self driving cars because they have all the info available and don't panic. Th…

> Some situations cant be avoided like if something falls off the car in front of you.

It can be avoided by allowing adequate breaking distance between the car in front. Most of the time having to make any decision between swerve and break can be avoided and someone that chooses swerve and kills someone should be charged with manslaughter.

Something non of the self driving cars seem to do is slow down for road conditions.

Re: Moral Machine

#39

Post by xvedejas »

In modern society we already have a mechanism for preventing machines from doing immoral things: hold someone accountable for its actions. It can be the owner or the manufacturer, or some mix of both, just someone should be responsible. This sets up the incentives for either better management by the owner, or for the manufacturer to design the car to avoid illegal harm. Why have a parallel system of morals aside from the law when we can just apply the law?

Re: Moral Machine

#40

Post by polkapolka »

joe_the_user wrote: post #37
whyte_mackay wrote: post #30

Earlier quoted context omitted.

I hear this often (that trolley problem is not relevant), but then I discovered that a lot of realistic ML fairness problems can be restated as a trolley problem. You have a classifier for credit assignment (giving a loan, etc.). The classifier is 99% accurate on the entire population. The classifier is 55% accurate on a small minority. You can improve the minority accuracy to 90% at the cost of 0.3% decrease of gene…

You have a classifier for credit assignment (giving a loan, etc.). The classifier is 99% accurate on the entire population. The classifier is 55% accurate on a small minority. Uh, concerning this hypothetical. I'll admit a scenario of this sort sounds appealing at first blush. But "99%" accuracy rate with credit assignment is transparently absurd if considers it for a second. There is a clear, significant limit to th…

You are thinking too US-centric here. There are jurisdictions where this is allowed.

Also don't stare yourself blind on the numbers.

Do you unfairly deny 5 minorities or erroneously deny 10 genpop?

This is reality for today's data scientists.

Right now in the US: You find an informative variable that acts as a proxy for race (such as Facebook likes). It is not forbidden by law to use it, but you have the data showing the proxy effect. Will you add it to the CTR model and get a raise, or do you act and speak up?

Will you let the trolley follow its course and run over 10.000 Chinese dissidents, or make the switch and run over 100 of your colleagues (and friends) who would benefit from a China expansion?

Post reply on HN