How do you teach a computer that correlation does not imply causation?

[Persevero]

Hey it's been a while since I've posted something on these forums, hope all of you have a good New Year's!

One of the things I've been doing rather than posting here is studying for one of the hardest exams in my undergraduate course, Econometrics.

One of the past exams I've been doing has this exercise which reminded me of one of Grumpy Cat's posts in her statistics thread, which has to do with the thread subject: The exam had a regression model which was trying to explain the number of crimes in a city with the following variables: expenditure on public safety, number of officers on the field, unemployment rate and a binary variable that was 1 if the majority of the population was under the age of 18.

It also gave the regression's results, and the funniest part was that the coefficients were saying the more a city spends on public safety and the more cops it has on patrol, the greater the number of crimes. To a human this makes sense after a little contemplation: What's really going up is the number of crimes being reported and handled by law enforcement, people aren't deciding to commit more crimes because of their being more police officers. It's impossible to measure the number of crimes that aren't reported.

In Grumpy Cat's case there were two completely unrelated things: Deaths by falling off beds and consumption of cheese . Can't remember the details right now because I can't find the thread.

Anyway more and more of society's management is being automated, and there are many science fiction depictions of the future which have a super computer running the whole thing - what I want to know is how is it possible to teach an AI that data isn't the end-all be-all? How do you teach an AI to weigh data based on its context, and which direction the implications go?

 

[Ste11aeres]

Ha, I'd like to start by teaching this to more human beings!

 

[Judge]

Anyway more and more of society's management is being automated, and there are many science fiction depictions of the future which have a super computer running the whole thing - what I want to know is how is it possible to teach an AI that data isn't the end-all be-all? How do you teach an AI to weigh data based on its context, and which direction the implications go?

Well, as long as something is within the realm of science fiction, I suppose the authors remain free to creatively put the cart in front of the horse.

 

[Ste11aeres]

The first question I want to know is, what exactly is AI?

I think that if AI means true real intelligence, than the ability to distinguish between causation and correlation will already be included as a part of such intelligence.

 

[Persevero]

Well there's the Turing test, which certainly needs the machine to know how to handle context, but to do that you don't need to distinguish correlation from causation, just data from previous parts of the conversation

 

[danielcollins]

You can program a computer to doo whatever you want it to. If you can build clever dependency trees and tables then logic around it, the computer will do it.

There is no vision though, or real truly independent thought - only that which is derived from programming. If you give someone or something the tools to work something out, does it then become intelligent? Or just deducing..

Proper hardcore thinking!

 

[Meistersinger]

The one thing you have to remember about a computer is that it is an intelligent idiot. It will do only what you tell it to do. If you tell it to spew out garbage for output, that's exactly what you get!

 

[Bad Wolf]

You can program a computer to doo whatever you want it to. If you can build clever dependency trees and tables then logic around it, the computer will do it.

There is no vision though, or real truly independent thought - only that which is derived from programming. If you give someone or something the tools to work something out, does it then become intelligent? Or just deducing..

Proper hardcore thinking!

As a simple proof that computer "intelligence" is really only mechanistic, consider trying to program a computer to "understand" natural language. When I see a computer correctly parse the following two statements, I might consider that computers are moving forward in having (or at least simulating) human-style intelligence:

Time flies like an arrow.

Fruit flies like a banana.

Until then, computer "intelligence" is merely simulated by a machine that is inherently stupid (it can execute in sequence, iterate over a loop, and move data back and forth in memory and registers), but creates the illusion of "intelligence" by being able to execute simple billions of simple operations incredibly quickly.

 

[Bad Wolf]

The one thing you have to remember about a computer is that it is an intelligent idiot. It will do only what you tell it to do. If you tell it to spew out garbage for output, that's exactly what you get!

But, remember, what you tell it and what you think you are telling it may be two entirely different things. Most computer anomalies or errors arise not because of a mistaken instruction, but because the programmer didn't really analyze and devise solutions for the essential core of the problem s/he was trying to solve. A computer's behavior is often human misconception writ large.

 

[Jaywalker]

It also gave the regression's results, and the funniest part was that the coefficients were saying the more a city spends on public safety and the more cops it has on patrol, the greater the number of crimes. To a human this makes sense after a little contemplation: What's really going up is the number of crimes being reported and handled by law enforcement, people aren't deciding to commit more crimes because of their being more police officers.

What's going on is that cities tend to spend more on public safety and hire more cops when crime increases. The data correlate in both direction (so the regression works in both directions), even though the causation works in only one direction.

As for determining if correlation implies causation, that's not something the software typically does. One tool statisticians use to reduce coincidental correlations is to use one dataset to build the regression model and a different dataset to test the model.

This method won't always work, though, particularly when a third variable influences both the dependent and independent variables. And it never detects incorrect direction of causation.

Sometimes random variables simply have similar distributions, so they will correlate even when they aren't otherwise related. (This is why it's easy to find datasets that spuriously correlate.)

Usually, when we do regression, we already know why one variable might affect another. We are interested in quantifying that effect and seeing if it is statistically significant.

 

[danielcollins]

But, remember, what you tell it and what you think you are telling it may be two entirely different things. Most computer anomalies or errors arise not because of a mistaken instruction, but because the programmer didn't really analyze and devise solutions for the essential core of the problem s/he was trying to solve. A computer's behavior is often human misconception writ large.

I agree, but what about the opposite argument (yes it's controversial) - if you teach someone the answer, does knowing the answer imply intelligence? Or is the question of intelligence in being able to apply rational logic and thought processes to deduce outcomes?

I've always wondered this. I work in an industry, where there are many with lots of textbook knowledge - few with applicable knowledge.

 

[StephF]

If the computer was to become sentient, it would realise that correlation doesn't equate to causation by itself.

 

[StephF]

if you teach someone the answer, does knowing the answer imply intelligence? Or is the question of intelligence in being able to apply rational logic and thought processes to deduce outcomes?

I would agree with the latter Daniel, simple possession of facts does not prima facie equate to intelligence. Intelligence is the extrapolation of existing facts to question further matters and make suppositions.
The desire to increase ones understanding and to further develop one's comprehension represents "intelligence" to me but then I am an aspie so I think differently from most of the mundanes I meet.

I have been told I come across as intelligent but arrogant and cold to some people but I guess thats kinda my Aspieness. I don't mean to look down on other people. I just choose to intereact with people on my own level.

 

[danielcollins]

I have been told I come across as intelligent but arrogant and cold to some people but I guess thats kinda my Aspieness. I don't mean to look down on other people. I just choose to intereact with people on my own level.

I'm the same, like you say it's an Aspie thing.. sometimes where sensitivity is needed - truth is delivered haha

 

[Tia Maria]

Agreed as well, true intelligence is being able to apply rational logic and thought processes to deduce outcomes. Not just with direct equations, but intangible analysis & judgement making as well.

And AI is not so science fiction although WHEN is the question. IMHO. Stay frosty you crazy chickens ;P

A.I. researchers say Elon Musk's fears 'not completely crazy' | Computerworld

 

[StephF]

Buk Buk Bukkaaaa !!