5 ms·

The higher level question is how should we respond when machines, trained on facts and data, systematically make decisions that go against our sensitivities?

halflings 9y ago
Your model can still be disfunctional even if it is trainer on "facts and data". No model has 100 accuracy, and it's easy to make a mistake when modeling and end up with unexpected issues (for example image recognition systems not working well with people of colour because of either a bias in the dataset, or non-robust methods)
to_bpr 9y ago
If the current climate is anything to base predictions on, I would put my money on the answer lying in hard-coding them to adhere to that which is deemed politically correct.
- creaghpatr 9y ago
  Thereby opening a market edge to algorithms that leverage politically incorrect hidden truths.
23443463453 9y ago
It's naive to think that there are objective facts and data that are sufficient to regulate human nature. Those "facts" and "data" you are referring to are absolutely meaningless, ameaning without the overlay of human intention. THAT's the crux of the terrifying future this tech community is racing towards, a future of "techtalitarianism", where emperor Zuckerberg, "elected" by use of his fraudulent wealth based on fraudulent network he uses to control people's minds, perception, and illusory "reality".
alvis 9y ago
Can't say our machines are trained based on facts, but purely data. What you get out of it is from what you feed into it. The challenge we are going to face in this AI era is that the training data can be biased or include some morally/politically incorrect information. Someone can intentionally manipulate the data and make us feel that the response from the machine is trustworthy and factual, which can be devastating if it is used by a dictating government. Mind the alternative facts!!!
- reader5000 9y ago
  This is not like a "fundamental" AI problem; just dont feed the AI shitty data.
  - oldsj 9y ago
    Right but what you call "shitty data" is subjective and can easily differ based on opinion.
    - reader5000 9y ago
      It's really an outdated concern anyway since most AIs run in realtime and are fed data direct from clickstreams / other realtime pipelines. It's unclear what a 'biased' data source means.
      - simias 9y ago
        Look at 4chan's successful "attacks" on chatbots: https://www.washingtonpost.com/news/the-intersect/wp/2016/03/24/the-internet-turned-tay-microsofts-fun-millennial-ai-bot-into-a-genocidal-maniac/ https://www.washingtonpost.com/news/the-intersect/wp/2016/03... You'll be hard pressed to find a single fact that would be considered so by the entire human population. There are people who are seriously disputing that the earth is round.
        zxcmx 9y ago
        I think that's a good example. Yes or no, is the world "round"? Well, it's sort of a very slightly lumpy oblate spheroid... it also ties into whether we agree on what "round" means. There's a knowledge problem involved - when answering a question you really have to understand the knowledge and intent of the speaker, establish a conversational frame. The first years of university are often spent learning certain "facts" and theories, which are technically incorrect (but useful), which will be torn down and shown as approximations later. Even things that look like "hard" facts are very often a matter of perspective.
kobeya 9y ago
garbage in, garbage out
TorKlingberg 9y ago
You cannot make decisions based purely on facts and data. You also need goals, values, preferences or what you prefer to call it.
- dogruck 9y ago
  I agree with you. I just don't know how it will work. Suppose we have a program that approves or denies loan applications. We input the acceptable default rate, and it makes the approval decisions. Then, we notice that the software disproportionately denies applications from a few demographics. What's the right reaction? Or, what if the software makes hiring decisions, where the objective function is "hire people that will, in aggregate, maximize the corporate stock price."
  - alvis 9y ago
    In fact, it is a new problem in fin tech. While ML can mimic human decision quite accurately nowadays, machines remain largely incapable to handle new information. Now suppose again we have a machine that rate load applications. In the past when you write your occupation as AI scientist, probably it won't make a difference than others. However, same occupation in the past few years, even you haven't got the default data to retrain the machine, a human would know the risk is significantly lesser than the majority. The problem? While we think the machine is at inferring future outcome, it is not in some important cases which we are largely unaware of.

halflings 9y ago

Your model can still be disfunctional even if it is trainer on "facts and data". No model has 100 accuracy, and it's easy to make a mistake when modeling and end up with unexpected issues (for example image recognition systems not working well with people of colour because of either a bias in the dataset, or non-robust methods)

to_bpr 9y ago

If the current climate is anything to base predictions on, I would put my money on the answer lying in hard-coding them to adhere to that which is deemed politically correct.

creaghpatr 9y ago
Thereby opening a market edge to algorithms that leverage politically incorrect hidden truths.

23443463453 9y ago

It's naive to think that there are objective facts and data that are sufficient to regulate human nature. Those "facts" and "data" you are referring to are absolutely meaningless, ameaning without the overlay of human intention. THAT's the crux of the terrifying future this tech community is racing towards, a future of "techtalitarianism", where emperor Zuckerberg, "elected" by use of his fraudulent wealth based on fraudulent network he uses to control people's minds, perception, and illusory "reality".

alvis 9y ago

Can't say our machines are trained based on facts, but purely data. What you get out of it is from what you feed into it. The challenge we are going to face in this AI era is that the training data can be biased or include some morally/politically incorrect information. Someone can intentionally manipulate the data and make us feel that the response from the machine is trustworthy and factual, which can be devastating if it is used by a dictating government. Mind the alternative facts!!!

reader5000 9y ago
This is not like a "fundamental" AI problem; just dont feed the AI shitty data.
- oldsj 9y ago
  Right but what you call "shitty data" is subjective and can easily differ based on opinion.
  - reader5000 9y ago
    It's really an outdated concern anyway since most AIs run in realtime and are fed data direct from clickstreams / other realtime pipelines. It's unclear what a 'biased' data source means.
    - simias 9y ago
      Look at 4chan's successful "attacks" on chatbots: https://www.washingtonpost.com/news/the-intersect/wp/2016/03/24/the-internet-turned-tay-microsofts-fun-millennial-ai-bot-into-a-genocidal-maniac/ https://www.washingtonpost.com/news/the-intersect/wp/2016/03... You'll be hard pressed to find a single fact that would be considered so by the entire human population. There are people who are seriously disputing that the earth is round.
      - zxcmx 9y ago
        I think that's a good example. Yes or no, is the world "round"? Well, it's sort of a very slightly lumpy oblate spheroid... it also ties into whether we agree on what "round" means. There's a knowledge problem involved - when answering a question you really have to understand the knowledge and intent of the speaker, establish a conversational frame. The first years of university are often spent learning certain "facts" and theories, which are technically incorrect (but useful), which will be torn down and shown as approximations later. Even things that look like "hard" facts are very often a matter of perspective.

kobeya 9y ago

garbage in, garbage out

TorKlingberg 9y ago

You cannot make decisions based purely on facts and data. You also need goals, values, preferences or what you prefer to call it.

dogruck 9y ago
I agree with you. I just don't know how it will work. Suppose we have a program that approves or denies loan applications. We input the acceptable default rate, and it makes the approval decisions. Then, we notice that the software disproportionately denies applications from a few demographics. What's the right reaction? Or, what if the software makes hiring decisions, where the objective function is "hire people that will, in aggregate, maximize the corporate stock price."
- alvis 9y ago
  In fact, it is a new problem in fin tech. While ML can mimic human decision quite accurately nowadays, machines remain largely incapable to handle new information. Now suppose again we have a machine that rate load applications. In the past when you write your occupation as AI scientist, probably it won't make a difference than others. However, same occupation in the past few years, even you haven't got the default data to retrain the machine, a human would know the risk is significantly lesser than the majority. The problem? While we think the machine is at inferring future outcome, it is not in some important cases which we are largely unaware of.