We take it fairly serious in the public sector in Denmark. We’re doing our own POCs and we’re inviting partners to build things with us. It’s very much a hit’n’miss field.
It works with recognition. It helps us sort through millions of page files looking for missing documents, both faster and with higher hit rate than using people. It helps us determine how much water is in drone photographed land areas based on the colours of the vegetation. Stuff like that.
Where it absolutely doesn’t work is for prediction and analysis.
Well, it’s not that it doesn’t work in BI as such. I kinda does, but we’ve been working with analysis and BI for 25-50 years. IBM wanted to sell us Watson analytics for BI as an example. We let them run their magic on some of our non-sensitive datasets and they came up with a range of interesting BI models and metrics. Every one of them were far, far inferior to our current human build BI setup, however, and a subscription to them would cost us around three full time analysts.
Prediction is much worse. The probability just isn’t there to use it for anything that isn’t as harmless as advertising. Worse than that, almost every model turns out biased.
So I guess you can say that I agree with you. It’s mostly just hype, but those parts that aren’t hype are incredibly useful.
I'm not going to say that ML is a panacea or always perfect but its also not one "thing". It's a complex field with a lot of required knowledge and ways to set things up.
IBM is largely to blame for that misunderstanding since they brand all of their offerings "Watson" and describe it as if it's one system; so if their system doesn't work, maybe none will. In reality, it's dozens of pre-implemented algorithms with varying degrees of quality. Importantly, how the data interacts with those algorithms requires domain expertise for the data and algorithms so it can be hit-or-mess if that expertise is not properly applied.
Bad headline. The author discusses why it would be a terrible idea to use inherently flawed ML approaches to establishing statistical models of criminal risk, and how opposed such a system is to the fundamental principles of the US justice system.
But that doesn’t mean the talk about using ML to cut human empathy out of more and more of the justice system is just “hype”. There’s no substance to the techniques proposed, that’s true. They are more dangerously biased than the humans they are meant to replace. But that doesn’t mean that these systems aren’t going to be purchased and implemented if we don’t push back harder against the general AI hype that pervades VC tech culture in general.
ML can recommend a movie to watch or a restaurant nearby, and when it fails terribly and reinforces biases and misunderstandings like it always does, the cost to society is minimal. But when you put people’s lives in its hands, the consequences of the inevitable failures are far more damaging.
>ML can recommend a movie to watch or a restaurant nearby, and when it fails terribly and reinforces biases and misunderstandings like it always does, the cost to society is minimal
I'd go further and say we should pay a lot of attention to this too. The subtle ways in which recommendation engines change what cultural goods we consume, including journalism, might very well have some pretty significant effects.
But yes, to essentially beta test rudimentary statistical models in actual court rooms or police stations were people's legal rights are at stake is pretty much insane.
Are there any cases of machine learning solidly outperforming other options in real-world applications? Is there anything it does better than a person? You can point at something like evolutionary algorithms and say "this antenna is better for this use than anything designed by hand" but can you do that with machine learning?
Evolutionary algorithms are a type of machine learning, so yes. Machine learned vision, speech, NLP, etc. outperform heuristic methods and those are just the recent advances.
A random forest can easily outperform a hand-made decision tree—and this is the important part—if it is set up correctly and you use a sufficient quantity of high quality data. Linear models have also been used for decades because they work.
I've heard that computer vision has huge advantages over humans in grading agricultural products, but from the brief discussions I've had I'm not sure how much of that is actually based on machine learning as opposed to other techniques (heuristics, ad-hoc algorithms, old-school statistical methods, etc.).
Jurisprudence is a rather theoretical study of the law, which does not seem well-suited to machine learning.
If, instead, you are thinking of learning from case law, let's consider patents. There is a lot of case law to be used in training from the Eastern District of Texas...
On paper, a ML approach to calculating risk of recidivism can make sense if you consider it as a silicon-based approximation of the criminal brain in making totally rational self-serving responses.
One caveat: even if you have a perfect ML system... never forget the 'human factor' when using these tools. A single fat-finger by a clerk in San Francisco lead to the release of a defendant who went on to commit murder in 2017. https://abc7news.com/details-of-mistaken-release-of-sf-twin-...
Since machine learning allows you to codify the bias of the system maybe the business to be built on top of judicial system is something that helps individual citizens instead of companies - use cases:
1. Tell you the chance you or members of your family will have a criminal case. Early warning system - do I want to leave this area?
2. Tell you what your chances are of conviction/time you will have to serve if convicted once you have a case.
Once it gets into tax law it might be really interesting.
It would be good if you could do stuff like give it an itinerary - I am going fishing this weekend with my awful in-laws - and it could give you sort of a list of things you might end up being charged with to look out for.
This seems to me to be a good article on the subject. It covers the bases soberly.
> The legitimacy of the judicial system is based on its status as a well-tested, publicly examinable, and heavily scrutinized tradition of practice. A shift to an unaccountable, opaque software system throws that basis away.
To me this is the fundamental problem with ML poisoning judicial systems: For better or for worse, we are abdicating our responsibility. Justice ain't easy. Automation should free up our time for more important (less automatable) tasks like justice and politics.
In so far as the rule of law means ensuring that actions have predictable consequences, I view it as imperative that we automate as much of the legal system as is technically possible (not saying we're there yet). In the US, federal law is growing increasingly complex, not to mention the added uncertainty introduced by judicial activism (judges reinterpreting old laws, ostensibly to make them fit modern society). As a result, the legal system is completely backlogged and unpredictable, requiring people to expend a lot of money retaining top lawyers who know how to game the system. If machine learning can help make the system more predictable, it could, on balance, improve the rule of law and the spread of justice, even if in a small fraction of cases a decision is made that differs from what a judge would have decided.
20 comments
[ 4.4 ms ] story [ 55.6 ms ] threadIt works with recognition. It helps us sort through millions of page files looking for missing documents, both faster and with higher hit rate than using people. It helps us determine how much water is in drone photographed land areas based on the colours of the vegetation. Stuff like that.
Where it absolutely doesn’t work is for prediction and analysis.
Well, it’s not that it doesn’t work in BI as such. I kinda does, but we’ve been working with analysis and BI for 25-50 years. IBM wanted to sell us Watson analytics for BI as an example. We let them run their magic on some of our non-sensitive datasets and they came up with a range of interesting BI models and metrics. Every one of them were far, far inferior to our current human build BI setup, however, and a subscription to them would cost us around three full time analysts.
Prediction is much worse. The probability just isn’t there to use it for anything that isn’t as harmless as advertising. Worse than that, almost every model turns out biased.
So I guess you can say that I agree with you. It’s mostly just hype, but those parts that aren’t hype are incredibly useful.
IBM is largely to blame for that misunderstanding since they brand all of their offerings "Watson" and describe it as if it's one system; so if their system doesn't work, maybe none will. In reality, it's dozens of pre-implemented algorithms with varying degrees of quality. Importantly, how the data interacts with those algorithms requires domain expertise for the data and algorithms so it can be hit-or-mess if that expertise is not properly applied.
But that doesn’t mean the talk about using ML to cut human empathy out of more and more of the justice system is just “hype”. There’s no substance to the techniques proposed, that’s true. They are more dangerously biased than the humans they are meant to replace. But that doesn’t mean that these systems aren’t going to be purchased and implemented if we don’t push back harder against the general AI hype that pervades VC tech culture in general.
ML can recommend a movie to watch or a restaurant nearby, and when it fails terribly and reinforces biases and misunderstandings like it always does, the cost to society is minimal. But when you put people’s lives in its hands, the consequences of the inevitable failures are far more damaging.
I forgot the link and am on mobile, but one can probably find it online.
Keep in mind that just because an algorithm e.g. sentences a black person to prison, it doesn't mean the algorithm is biased.
I'd go further and say we should pay a lot of attention to this too. The subtle ways in which recommendation engines change what cultural goods we consume, including journalism, might very well have some pretty significant effects.
But yes, to essentially beta test rudimentary statistical models in actual court rooms or police stations were people's legal rights are at stake is pretty much insane.
A random forest can easily outperform a hand-made decision tree—and this is the important part—if it is set up correctly and you use a sufficient quantity of high quality data. Linear models have also been used for decades because they work.
Civil cases could benefit enormously from ML, helping judges with rulings suggestions based on jurisprudence.
What makes you think “ML” would even be able to do that?
If, instead, you are thinking of learning from case law, let's consider patents. There is a lot of case law to be used in training from the Eastern District of Texas...
On paper, a ML approach to calculating risk of recidivism can make sense if you consider it as a silicon-based approximation of the criminal brain in making totally rational self-serving responses.
One caveat: even if you have a perfect ML system... never forget the 'human factor' when using these tools. A single fat-finger by a clerk in San Francisco lead to the release of a defendant who went on to commit murder in 2017. https://abc7news.com/details-of-mistaken-release-of-sf-twin-...
1. Tell you the chance you or members of your family will have a criminal case. Early warning system - do I want to leave this area?
2. Tell you what your chances are of conviction/time you will have to serve if convicted once you have a case.
Once it gets into tax law it might be really interesting.
> The legitimacy of the judicial system is based on its status as a well-tested, publicly examinable, and heavily scrutinized tradition of practice. A shift to an unaccountable, opaque software system throws that basis away.
To me this is the fundamental problem with ML poisoning judicial systems: For better or for worse, we are abdicating our responsibility. Justice ain't easy. Automation should free up our time for more important (less automatable) tasks like justice and politics.