AI Has a Hallucination Problem That's Proving Tough to Fix

Discussion in 'other software & services' started by ronjor, Mar 11, 2018.

  1. ronjor

    ronjor Global Moderator

    Joined:
    Jul 21, 2003
    Posts:
    162,650
    Location:
    Texas
    Tom Simonite 03.09.18
     
  2. 142395

    142395 Guest

    Not just that. Seeing things is more dynamic and active process than thought in past.
    Only safe way is not to use ML for anything vital which can cause, literary, death etc.. But it seems majority of people are too optimistic.
     
  3. RockLobster

    RockLobster Registered Member

    Joined:
    Nov 8, 2007
    Posts:
    1,812
    Because they have $$$ in their eyes.
     
  4. 142395

    142395 Guest

    Yes, they know who succeeded in this will make big business. Well, I might be a bit harsh as I can't say we should forbid airplane as it can fall. But while airplane just extend human's capability, AI is aimed at replacing a part of role of human. The problem is current 'AI' can't think like human, nor see like human, nor listen, etc., in the end it's Neumann computer. It should only be used in safe use such as Go game, or when the benefit clearly outweigh the risk such as high-place work.

    Every year we see articles that "Email was not designed with security in mind because in that era...", "DNS was not...", "USB...". But we're going to repeat the history so we'll hear 30y later "automated car was not designed with...". There have to be strict, thorough, comprehensive scrutiny and regulation, but who do this? Companies won't as it costs too much and also delays the development, meaning they'll be behind competitors. But politicians will think the same, if US put such a regulation, they'll be behind China. Only hope is we international citizens speak up together.
     
    Last edited by a moderator: Mar 14, 2018
  5. RockLobster

    RockLobster Registered Member

    Joined:
    Nov 8, 2007
    Posts:
    1,812
    @142395 You make some very good points but when its a battle between, what is the right thing to do vs $$$, you dont need to look very far to see which usually wins.
     
  6. Mrkvonic

    Mrkvonic Linux Systems Expert

    Joined:
    May 9, 2005
    Posts:
    10,207
    The more we simulate perceived reality with software, the more will it resemble our brains.
    Humans are prone to hallucinations - even subtle things like remembering things not quite as they were.
    So the more we humanize AI, the more it will be like us.
    Mrk
     
  7. 142395

    142395 Guest

    If nobody wanted automated car, there's no $$$ for developing it.;) But I know, those who're happy with fingerprint scanner and Google home would have no reason to object to automated car or 'smart' sth.:(
     
  8. 142395

    142395 Guest

    As the article mentions, NN by no means resemble our brain. 'Hallucination' in this case is just a common catchy word often used by these media. Hallucination and sensory illusion actually is a byproduct of adaptation. Only recently it began to be understood in that context thanks to developments of mathematical models in neuroscience so now we have some control to intentionally make many kinds of illusion. Even real hallucination in schizophrenia can be reasonably understood in that context with Bayesian model. Generally ML don't bother to adapt to environment to survive, so they don't have human like hallucination or illusion, and what the article described is more of a data poisoning or unintentional behavior which result from that we have nearly no control on precise work of NN. The fundamental difference is ML just passively process visuals while we see things actively, making philosophical question what the 'time' or even 'real' is for us. Well, if I speak too much about it I'll endanger my real identity lol. If you have interest but don't want to be bothered by math, search for 'postdiction' with some other terms such as 'illusion'.

    You know, NN is far from actual work by neurons. It's possible and already done to make human-like neural system, as we have plenty of differential eq models to represent neural activity. But it seems mostly they are studied by neuroscientist and not by ML guys, probably because it's too complex and they don't expect that to be usable for production purpose.
     
  9. Marcelo

    Marcelo Registered Member

    Joined:
    Oct 11, 2005
    Posts:
    274
    Location:
    Rio de Janeiro, Brazil.
    Let me say I will be the first one in line to buy an automated car. I absolutely hate driving and sometimes public transportation is not an option neither is hiring a driver. If you are worried someone will be tracking you through your car there are a lot of simpler ways of doing that that are already available...
     
  10. Mrkvonic

    Mrkvonic Linux Systems Expert

    Joined:
    May 9, 2005
    Posts:
    10,207
    Yuki - it does not resemble - not yet.

    I wrote the more we ... it does not mean it's similar - but it will become so over time (10s, 100s of years).

    And we need not get hung up on terminology, but eventually, AI may become aware of its environment, at which point, the survival question will arise.

    Once that happens - and every time the environment changes - the algorithms may change as well, leading to different results in what the software "thinks" hence the deterministic lie. The word hallucination is definitely hypy. though.

    Mrk
     
  11. 142395

    142395 Guest

    IDK if you say that after reading the article, besides many other related articles on the web (even in this forum only, there're many posts already). Tracking is actually a tiny part of problems. More serious is these smart sth have to be connected to network and it means remote hackers can exploit them, there're already cases (tho mostly PoC) and if it was car, it directly causes your death. It'll also come true a terrorist hacks 100s of cars at once, and as mentioned in the article, attacker don't necessarily need to attack your car from internet, he just makes a trick in a road sign and all automated car which see that may crash. You may argue driving yourself can also cause death by accident, but you know that depends mostly on yourself (sometimes not you but another driver behind your car, but that is also true to automated car so we can ignore that from consideration) and established technology of modern cars. OTOH, ML is far from established tech, and worse, as discussed in the articles, solving the problem fundamentally is very hard thanks to the nature of NN. It's just a matter of time to hear about news of death or even catastrophe by these (not at all intelligent) 'AI's. I know I don't put my life on it, just like I don't on a drunk driver.
     
  12. 142395

    142395 Guest

    Yes.:)
    Some ppl are actually trying to make real AI, but it will take at least several decades to come true. We still know almost nothing about how actually human thinks.
     
  13. Mrkvonic

    Mrkvonic Linux Systems Expert

    Joined:
    May 9, 2005
    Posts:
    10,207
    We will discover that true software and hardware. I believe the technology is an abstraction of our brain. On a macro scale, the global internet grid is a primitive synaptic map, albeit one without real (self) awareness. It needs another 6-7 orders of magnitude before it can begin computing human-complex tasks, and that's when the science will get interesting.
    Mrk
     
  14. Marcelo

    Marcelo Registered Member

    Joined:
    Oct 11, 2005
    Posts:
    274
    Location:
    Rio de Janeiro, Brazil.
    I did and it does not worry me. Every technology ever developed was once impossible. Whatever you are using now to read this answer was impossible a few decades ago. If this technology reaches the public it will be quite polished.
     
  15. JoWazzoo

    JoWazzoo Registered Member

    Joined:
    Nov 11, 2008
    Posts:
    241
    Location:
    Ether
    Interesting stuff indeed. I was involved as a Partner in a Limited Partnership (Knowledge Application Software) and one of our investments was in Knowledge Engineering - basically what is today viewed as AI. Heady stuff indeed. That was 30 years ago - very cutting edge. Just wish I had put more money in. :) And it took us 12 years to sell out.
     
  1. This site uses cookies to help personalise content, tailor your experience and to keep you logged in if you register.
    By continuing to use this site, you are consenting to our use of cookies.