February 23, 2018 at 8:14 am
xsevensinzx - Thursday, February 22, 2018 3:50 PMWe do. I can run all different types of checks on the results and so forth in terms of measuring errors as I'm sure you know as you've been playing with R. But these are not the only things you can go on with explaining the output. This is where the math and statistics comes into play much like it does for students having to show their work.
True, but there's also the error/risk assessment that takes place. A human has to learn to judge what risk means, or what a change means. Is 55% much greater than 45%? Some people might think these aren't that different as every 1% increase in one side can mean a 1% decrease in the other, so a 1% change looks like 2. However, if a machine gives you an output, your view might change. You might drastically view this as significant, or very insignificant.
We, as humans, have to often still make judgements on what results from a score/prediction/analysis mean. At least when we're building software around these items. It's good to know how the result is arrived at, or be able to defend it, once we put some other decision or continued workflow on top of the result. I know some data scientists work hard to understand and ensure the results are accurate in terms of what they're studying. They know the math and stats. What I worry about is people like me that aren't strong in stats taking these values and using them without deep understanding. That's where I worry we'll see lots of AI/ML implementations that are severely flawed. Especially when we don't have tools for the lesser skilled to understand how/why an output appears.
February 23, 2018 at 8:15 am
David.Poole - Thursday, February 22, 2018 11:52 PMI'm cursing Oliver Cromwell. 350 years in his grave and his taking away the right to say "It's witchcraft" comes back to haunt.
I would hope this isn't the case, but I can see lawyers/solicitors using this as a defense. "This algorithm is a black box", when sued because of some action taken.
February 23, 2018 at 8:34 am
Steve Jones - SSC Editor - Friday, February 23, 2018 8:15 AMI would hope this isn't the case, but I can see lawyers/solicitors using this as a defense. "This algorithm is a black box", when sued because of some action taken.
The thing is, that to be functional 'AI' it needs to go beyond definable algorithms (we've used them for years). How would you defend or attack the result of an AI decision based on complex weightings assigned dynamically over millions of training samples, as well as feedback from daily experience.
You basically can't. Machine learning in the real world is fluid, and our legal system has no way to accommodate that.
...
-- FORTRAN manual for Xerox Computers --
February 23, 2018 at 9:10 am
Hugo Kornelis - Thursday, February 22, 2018 4:32 AM@peter-2: ML is now beyond the point of doing broad analyis and coming up with customer groups that are good targets for the next campaign. More and more applications train a model based on generic data and then use that for individual predictions. E.g., you aply for a healthcare policy, provide some data, and then the AI determines that you are 65% likely to be an increased risk so you have to pay more.or even get refused. And yes, I can certainly see how customers would ask for an explanation.@steve-2: When reading your editorial I actually get concerned over an issue that is broader than GDPR. That concern is to an extent embedded in my response to Peter. Apparently, we are now using tool where even the people that decide to implement those tools and change the business process to use them do not really understand how the results are computed. And yet we use these results and allow them to have high impact on our business decisions. Law enforcement or border control use AI and ML algorithms to decide who gets screened and who can pass. Insurance companies use this data to determine who is accepted for a policy. Perhaps one day (perhaps now already though I don't think so) doctors will use these algoruthms to determine which patient will be the recipient for a kidney that has become available. If nobody can explain how those results are computed, then nobody can verify the results. So why should we trust them? What are we building our future on?
Hugo I agree with your concern. To add to them:
- there is no litmus test to define "generic data". It would be very difficult right now to know whether the ML algorithm is flawed/biased or simply reflecting the noise (whether it's aleatory or introduced purposefully to skew the model) in the underlying data sample. You'd almost have to have a customer auditable way to review said data set since it too drives how the AI decision affects their outcomes.
- how do you train an algorithm to "be ethical"? There are lots of cases where we as societies have decided to act against what raw numbers might tell us, simply because it's the right thing. Health care comes to mind: politics aside - handling the health care for a population defies normal market-driven rules, since it makes no economic sense to ever cover the very sickest among us (i.e. those who need it the most).
----------------------------------------------------------------------------------
Your lack of planning does not constitute an emergency on my part...unless you're my manager...or a director and above...or a really loud-spoken end-user..All right - what was my emergency again?
February 23, 2018 at 11:56 am
jay-h - Friday, February 23, 2018 8:34 AMSteve Jones - SSC Editor - Friday, February 23, 2018 8:15 AMI would hope this isn't the case, but I can see lawyers/solicitors using this as a defense. "This algorithm is a black box", when sued because of some action taken.The thing is, that to be functional 'AI' it needs to go beyond definable algorithms (we've used them for years). How would you defend or attack the result of an AI decision based on complex weightings assigned dynamically over millions of training samples, as well as feedback from daily experience.
You basically can't. Machine learning in the real world is fluid, and our legal system has no way to accommodate that.
And the training has to be really really good too, case in point: http://rocknrollnerd.github.io/ml/2015/05/27/leopard-sofa.html
February 23, 2018 at 12:05 pm
jay-h - Friday, February 23, 2018 8:34 AMSteve Jones - SSC Editor - Friday, February 23, 2018 8:15 AMI would hope this isn't the case, but I can see lawyers/solicitors using this as a defense. "This algorithm is a black box", when sued because of some action taken.The thing is, that to be functional 'AI' it needs to go beyond definable algorithms (we've used them for years). How would you defend or attack the result of an AI decision based on complex weightings assigned dynamically over millions of training samples, as well as feedback from daily experience.
You basically can't. Machine learning in the real world is fluid, and our legal system has no way to accommodate that.
So we deny benefits or employment or insurance to someone based on some algorithm no one can describe to any level of accuracy, and the answer back to the consumer is essentially "just trust me - you can't have this <insert description>"? How did we ever get to the point where we thought we could make decisions like this and support them?
Here's one outcome I can predict - this will end up in front of the supreme court sooner or later.
----------------------------------------------------------------------------------
Your lack of planning does not constitute an emergency on my part...unless you're my manager...or a director and above...or a really loud-spoken end-user..All right - what was my emergency again?
February 23, 2018 at 12:28 pm
Remember the legal systems between the U.S. and Europe vary significantly.
Lawyers in this country work by bringing lawsuits against companies.
The law in Europe works more via regulation, that is the government taking action against companies.
412-977-3526 call/text
February 23, 2018 at 12:58 pm
patrickmcginnis59 10839 - Friday, February 23, 2018 11:56 AMAnd the training has to be really really good too, case in point: http://rocknrollnerd.github.io/ml/2015/05/27/leopard-sofa.html
Ah yes.
By calling it 'intelligence', I think AI is being much oversold. It's a powerful tool, but it's really primarily correlation, which is a very low level of intelligence. It's the way animals learn. An animal may sense from experience that certain things in the environment means rain is approaching, but has absolutely zero comprehension of meteorology.
Correlation is only the first, and simplest step in human intelligence. When we observe things together, the next question is why, and how the two might be connected. Is one a cause of the other. We can create a plan to test, to see if A causes B or B causes A, or neither. Human intelligence involves mentally modelling the different possible interpretations to determine what, if anything, makes sense. Human intelligence whether if X is true can we expect Y to be true also. And if not, why not and what does that imply. Machine learning does not do this. Machine learning does not differentiate cause and effect. It does not say (as we must sometimes) 'this interpretation cannot be right, there must be something else at work'.
Machine learning does not look at the more global structure. The Google learning system came up with some conclusions that were humorously politically incorrect (women are homemakers) to others that were much worse.
On the legal front, machines also lack the concept of 'intent' which is very significant in our culture and legal system. But intent is critical, it helps humans to at times override the strictly mathematical analysis in the interest of outside the confines of the data set. Hence it's a significant part of our historic legal system. Artificial thinking is essentially indifferent, so it requires some strong human curbs... but since the systems are so complex, anticipating the curbs necessary is no small task, and large errors are just about certain to happen.
...
-- FORTRAN manual for Xerox Computers --
February 23, 2018 at 4:33 pm
Matt Miller (4) - Friday, February 23, 2018 12:05 PMSo we deny benefits or employment or insurance to someone based on some algorithm no one can describe to any level of accuracy, and the answer back to the consumer is essentially "just trust me - you can't have this <insert description>"? How did we ever get to the point where we thought we could make decisions like this and support them?
Here's one outcome I can predict - this will end up in front of the supreme court sooner or later.
Not taking that bet
February 23, 2018 at 4:37 pm
jay-h - Friday, February 23, 2018 12:58 PMAh yes.By calling it 'intelligence', I think AI is being much oversold. It's a powerful tool, but it's really primarily correlation, which is a very low level of intelligence. It's the way animals learn. An animal may sense from experience that certain things in the environment means rain is approaching, but has absolutely zero comprehension of meteorology.
Correlation is only the first, and simplest step in human intelligence. When we observe things together, the next question is why, and how the two might be connected. Is one a cause of the other. We can create a plan to test, to see if A causes B or B causes A, or neither. Human intelligence involves mentally modelling the different possible interpretations to determine what, if anything, makes sense. Human intelligence whether if X is true can we expect Y to be true also. And if not, why not and what does that imply. Machine learning does not do this. Machine learning does not differentiate cause and effect. It does not say (as we must sometimes) 'this interpretation cannot be right, there must be something else at work'.
...
I'll disagree here. The more I study that ML/AI does, the more I find that these systems are being taught the same way humans are. They aren't sentient by any stretch, nor are they making leaps we couldn't make. What they do is find correlation and complex patterns that are very difficult for most humans, or perhaps are those that we wouldn't take the time to find. There are good arguments that these systems are just learning to be faster than humans at interpreting large amounts of data, but not better.
What about the humorous issues? Well, some of this is that the humans don't really understand the complete complexity and end up not providing enough data. They think they are solving problem x, but don't realize there are other implications and issues. As a result, they don't have data to feed into the system about the other issues, or more likely, they haven't narrowly defined the problem enough. It seems that ML systems work best when the scope of what is being done is narrow.
February 23, 2018 at 9:10 pm
I'll just add that machine learning is learning from what is happening. It's essentially, optimizing itself. That way when the rain comes, it can learn what that means. The first time, it may do nothing and see what impact rain has on it. Then it may seek shelter and see the impact there. It's collecting data and acting on it depending on what it's been told to do (supervised) or what it wants to do (unsupervised).
What we are typically seeing, like with Google for example, the machine learning is referencing lots and lots of historical data. It's solving problems like when you type in a search engine a misspelling and picks up on similar misspellings for what it thinks you may be searching for. The more you say it's right or wrong, hopefully the smarter the optimization is the next time someone like you searches for something. If you ask me, it's been pretty good as of lately. It's even starting to stretch across products like detecting what you are searching for quicker on a few letters based on maybe what you already searched for on YouTube or something.
Another good one is Facebook. I'm almost certain it's measuring the time in seconds you hover over a post on your feed. Then it shows related posts to that content as you continue to scroll downward. For example, if you saw a friend post a video about the recent gun rights discussions, the machine learning algorithms bring those posts up to the forefront as you scroll down from other friends.
To end here, if you feel any of those two problems needs something equivalent to needing to know meteorology to avoid the rain, then I feel you are really overthinking the usage and the intent of ML in the real world. It's not trying to re-create a human brain in every application. It's trying to automate, learn from data, and of course, make smart decisions for particular subset of use cases -- NOT every use case or unknown use cases.
"You only had one job..."
February 26, 2018 at 8:51 am
Steve Jones - SSC Editor - Friday, February 23, 2018 4:37 PMI'll disagree here. The more I study that ML/AI does, the more I find that these systems are being taught the same way humans are. They aren't sentient by any stretch, nor are they making leaps we couldn't make. What they do is find correlation and complex patterns that are very difficult for most humans, or perhaps are those that we wouldn't take the time to find. There are good arguments that these systems are just learning to be faster than humans at interpreting large amounts of data, but not better.What about the humorous issues? Well, some of this is that the humans don't really understand the complete complexity and end up not providing enough data. They think they are solving problem x, but don't realize there are other implications and issues. As a result, they don't have data to feed into the system about the other issues, or more likely, they haven't narrowly defined the problem enough. It seems that ML systems work best when the scope of what is being done is narrow.
There is no doubt that a lot can be gained by processing power, but still, correlation to 'artificial intelligence' is comparable to power saws and nail guns to carpentry. Huge increase in efficiency but still at the basic level.
Statistical correlation is an automaton operation. Requires zero intelligence, just processing power. The step up, from simply grinding large amounts of information, to actually understanding that information, building and testing thought models to see what is plausible, to suddenly hit that moment where a bunch of data actually creates a moment of logical understanding... these are a huge jump in sophistication, and a jump that few, if any animals, have made. Even high horsepower correlation cannot distinguish cause and effect.
...
-- FORTRAN manual for Xerox Computers --
February 26, 2018 at 3:27 pm
Victor Kirkpatrick - Thursday, February 22, 2018 6:57 AMHugo Kornelis - Thursday, February 22, 2018 6:09 AMVictor Kirkpatrick - Thursday, February 22, 2018 5:38 AMGDPR = government overreach and over-regulation. Yes, we have to comply, and there are some aspects which make sense to protect consumers, but overall it's just too much, kind of like HIPAA in the USA. Only small parts are adopted and it simply canNOT be enforced without growing government even more than it already is (way too bloated). Sorry.Are you sure?
https://www.identityforce.com/blog/2017-data-breaches
https://gizmodo.com/the-great-data-breach-disasters-of-2017-1821582178
https://www.itgovernance.co.uk/blog/list-of-data-breaches-and-cyber-attacks-in-2017-33-8-million-records-leaked/Seems to me that at least currently there is insufficient regulation and insufficient incentive for businesses to get their act together and keep our personal details safe.
(Note that it took me less then a minute to find these links - I typed "data breaches in 2017" in Google and clicked the top 2 links. There are a lot more results on that search)(EDIT: Added one more link, the fifth search result, because it is an even longer list and because the author of that list expresses the hope that EU GDPR will lead to improvement in the near future - seems fiting for this discussion)
No one doubts that companies treating our data too loosely is a problem. This issue gets too political. Bottom line: if you believe government is the answer to your problems, you are for overreaching solutions such as GDPR. I just believe the regulation is too burdensome on business. In the case of a breach, that is when the government comes down on you with full force, and not before.
I would prefer elected officials writing the law on this instead of unaccountable bureaucrats. Exposing a business to serious financial liability if they were found in a court of law to be compromising an individual's privacy can have the same deterrence effect, especially in a class action suit. Trying to keep the regulations current as the technology evolves is also a serious hurdle for everyone. I think the more tightly it's defined the more likely it is to be ineffective.
February 27, 2018 at 11:22 am
I doubt this was the intention of the authors, though I do hope that this doesn't prevent the use of newer tools and technologies. What I'd like to see take place is more research and understanding into how the various algorithms we want to use for ML and AI technologies work, perhaps with some more detailed analysis of the inner workings of the models.
Cartoon showing Magic 8 Ball sitting on investment analyst's desk:
https://s3.amazonaws.com/lowres.cartoonstock.com/business-commerce-analytical_software-upgrades-magic_8_balls-magic_ball-8_ball-dcrn1512_low.jpg
"Do not seek to follow in the footsteps of the wise. Instead, seek what they sought." - Matsuo Basho
February 27, 2018 at 1:18 pm
HighPlainsDBA - Monday, February 26, 2018 3:27 PMVictor Kirkpatrick - Thursday, February 22, 2018 6:57 AMHugo Kornelis - Thursday, February 22, 2018 6:09 AMVictor Kirkpatrick - Thursday, February 22, 2018 5:38 AMGDPR = government overreach and over-regulation. Yes, we have to comply, and there are some aspects which make sense to protect consumers, but overall it's just too much, kind of like HIPAA in the USA. Only small parts are adopted and it simply canNOT be enforced without growing government even more than it already is (way too bloated). Sorry.Are you sure?
https://www.identityforce.com/blog/2017-data-breaches
https://gizmodo.com/the-great-data-breach-disasters-of-2017-1821582178
https://www.itgovernance.co.uk/blog/list-of-data-breaches-and-cyber-attacks-in-2017-33-8-million-records-leaked/Seems to me that at least currently there is insufficient regulation and insufficient incentive for businesses to get their act together and keep our personal details safe.
(Note that it took me less then a minute to find these links - I typed "data breaches in 2017" in Google and clicked the top 2 links. There are a lot more results on that search)(EDIT: Added one more link, the fifth search result, because it is an even longer list and because the author of that list expresses the hope that EU GDPR will lead to improvement in the near future - seems fiting for this discussion)
No one doubts that companies treating our data too loosely is a problem. This issue gets too political. Bottom line: if you believe government is the answer to your problems, you are for overreaching solutions such as GDPR. I just believe the regulation is too burdensome on business. In the case of a breach, that is when the government comes down on you with full force, and not before.
I would prefer elected officials writing the law on this instead of unaccountable bureaucrats. Exposing a business to serious financial liability if they were found in a court of law to be compromising an individual's privacy can have the same deterrence effect, especially in a class action suit. Trying to keep the regulations current as the technology evolves is also a serious hurdle for everyone. I think the more tightly it's defined the more likely it is to be ineffective.
Yes, most likely if it were up to the business they wouldn't bother with security. Ok next step in this line of argument is that people will say that "the free market will then reward businesses that DO bother with security" but then without government, the ENTIRE market segment could then collude to agree to let security lapse for everyone, and then the entire market segment could reap the benefits of reduced costs at the expense of consumer welfare.
I think whats needed is a calibrated measure of government oversight, but that's not going to happen in today's freakshow environment in the US for example.
Viewing 15 posts - 16 through 29 (of 29 total)
You must be logged in to reply to this topic. Login to reply
