3 ms·

Most of the recent advances don't really constitute what you are probably thinking of as "serious" AI. What's happening now is mostly just scientists putting ne

Most of the recent advances don't really constitute what you are probably thinking of as "serious" AI. What's happening now is mostly just scientists putting newly available massively parallel processing units (GPUs etc.) to work at sort of "brute forcing" the problem of AI. You could technically say it's a step above brute-forcing in that, after randomly assigning neural weights, gradient descent and other tricks are used to more quickly settle into good solutions. But the science is simply nowhere near the level of the human brain, and the current state of the art is still only "proficient" at relatively simple pattern matching, rather than coming up with decisions that actually influence the state of the world that it is perceiving. The deep architectural insights that will likely be needed before computers can approach the level of human intelligence still haven't been made yet, it seems to me based on what I know of the current research.

nl 12y ago
Yeah, this is almost completely wrong - or at least outdated thinking. While it's entirely likely that better AI algorithms will become available, the progress being made by "brute forcing" problems using lots of data and lots of computing power is much more than was expected 10 years ago. For example, Facebook's facial recognition software now correctly recognizes faces 97.25% of the time. Humans do 97.53%[1]. I'd encourage you to read the Norvig et. al paper: "The Unreasonable Effectiveness of Data" [1] http://www.technologyreview.com/news/525586/facebook-creates-software-that-matches-faces-almost-as-well-as-you-do/ http://www.technologyreview.com/news/525586/facebook-creates... [2] https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/35179.pdf https://static.googleusercontent.com/media/research.google.c...
- paganel 12y ago
  > For example, Facebook's facial recognition software now correctly recognizes faces 97.25% of the time. Humans do 97.53%[1]. One could say that is only data, nothing "intelligent" about pattern-matching it. But I'm biased, I'm a non-believer when it comes to reaching the singularity, as in I don't think we will ever be able to build something that will understand jokes, sexual innuendo or who/which will experience the Madeleine effect (http://en.wikipedia.org/wiki/Madeleine_%28cake%29#Literary_reference http://en.wikipedia.org/wiki/Madeleine_%28cake%29#Literary_r...)
  - skj 12y ago
    The difference between "artificial intelligence" and "automated problem solving" is how much the algorithm designers understand what's going on. Computers will certainly be able to understand jokes or innuendo. At least as well as people do, anyway. As far as the Madeleine effect goes, I'm not sure why you consider that intelligence?
    - paganel 12y ago
      > Computers will certainly be able to understand jokes or innuendo. At least as well as people do, anyway. Serious question, are we close to anywhere on that? I last looked at AI-related stuff around 10 years ago, when "sentiment analysis" as it was then called was only capable of approximating what "good press" and "bad press" meant. > As far as the Madeleine effect goes, I'm not sure why you consider that intelligence? I'm making a generalization in thinking that "intelligence" ultimately means "human intelligence", and I'm also assuming (probably wrong, there's no exact science for it) that things like the "Madeleine effect" sit at the core of us being humans, and the third (and last generalization) in thinking that if you're not experiencing things like the "Madeleine effect" you're not really human-like, so you cannot really be "intelligent". Writers like Asimov and Arthur C. Clarke have written about this problem (what it means to be human when "robots" will come about) way better than I could ever dream of doing, but my favorite presentation of the potential issues when the Singularity will come around remains Stanislaw Lem's "The Cyberiad". Or we could continue focusing on technicalities only and call it a day.
  - TheOtherHobbes 12y ago
    Intelligent behaviour in humans and other animals shows evidence of wide-scale situational awareness, and independent, unpredictable motivation. Before you use pattern recognition (i.e. pre-defined, remembered, or improvised abstraction for invariants) to solve a problem, you have to understand there's a problem to be solved. ML seems to be doing well at pattern recognition, but that second problem - understanding what needs to be done in a situation, and why, and accurately predicting possible developments and consequences - is possibly more critical for real singularity-style intelligence. It's machine reasoning as opposed to machine learning - in the street-wise sense, not the theorem proving sense.
- shock-value 12y ago
  Facial recognition is exactly the kind of "simple pattern matching" that I'm talking about. It's definitely useful, and definitely fits the label "artificial intelligence" (at least compared to other solved tasks), but it simply is just one minuscule piece of what actually constitutes human intelligence. Bear in mind that the original post referenced the "singularity" -- when artificial intelligence is smarter and more capable than actual human intelligence. To be clear, I'm not saying that new AI paradigms are needed to improve that facial recognition score from 97.25% to 99+% (the existing techniques might actually be optimal anyway). I'm saying they are needed to even be able to attempt other tasks that actual humans find effortless, such as interpreting literature, creating a song, carrying on a conversation, decorating a room in an appealing way, understanding symbolic math, having this debate with you, etc. (Certainly symbolic math software has been developed, but no artificial intelligence has been developed that "understands" it and can put it into larger context, particularly one based on neural networks a la the real human brain.) I don't think my thinking is outdated (or even that controversial). It's mostly informed by a Deep Learning class I recently took taught by Yann LeCun, who is definitely at the forefront of the field (although these opinions are entirely my own). EDIT: Also, when I mention new "AI paradigms", I do expect them to build on the deep neural networks currently being studied. But more sophisticated architectures than feedforward nets or even the simple recurrent nets that I've seen so far will be needed, I think. And that will likely require some breakthroughs in theory as well as engineering.

nl 12y ago

Yeah, this is almost completely wrong - or at least outdated thinking. While it's entirely likely that better AI algorithms will become available, the progress being made by "brute forcing" problems using lots of data and lots of computing power is much more than was expected 10 years ago. For example, Facebook's facial recognition software now correctly recognizes faces 97.25% of the time. Humans do 97.53%[1]. I'd encourage you to read the Norvig et. al paper: "The Unreasonable Effectiveness of Data" [1] http://www.technologyreview.com/news/525586/facebook-creates-software-that-matches-faces-almost-as-well-as-you-do/ http://www.technologyreview.com/news/525586/facebook-creates... [2] https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/35179.pdf https://static.googleusercontent.com/media/research.google.c...

paganel 12y ago
> For example, Facebook's facial recognition software now correctly recognizes faces 97.25% of the time. Humans do 97.53%[1]. One could say that is only data, nothing "intelligent" about pattern-matching it. But I'm biased, I'm a non-believer when it comes to reaching the singularity, as in I don't think we will ever be able to build something that will understand jokes, sexual innuendo or who/which will experience the Madeleine effect (http://en.wikipedia.org/wiki/Madeleine_%28cake%29#Literary_reference http://en.wikipedia.org/wiki/Madeleine_%28cake%29#Literary_r...)
- skj 12y ago
  The difference between "artificial intelligence" and "automated problem solving" is how much the algorithm designers understand what's going on. Computers will certainly be able to understand jokes or innuendo. At least as well as people do, anyway. As far as the Madeleine effect goes, I'm not sure why you consider that intelligence?
  - paganel 12y ago
    > Computers will certainly be able to understand jokes or innuendo. At least as well as people do, anyway. Serious question, are we close to anywhere on that? I last looked at AI-related stuff around 10 years ago, when "sentiment analysis" as it was then called was only capable of approximating what "good press" and "bad press" meant. > As far as the Madeleine effect goes, I'm not sure why you consider that intelligence? I'm making a generalization in thinking that "intelligence" ultimately means "human intelligence", and I'm also assuming (probably wrong, there's no exact science for it) that things like the "Madeleine effect" sit at the core of us being humans, and the third (and last generalization) in thinking that if you're not experiencing things like the "Madeleine effect" you're not really human-like, so you cannot really be "intelligent". Writers like Asimov and Arthur C. Clarke have written about this problem (what it means to be human when "robots" will come about) way better than I could ever dream of doing, but my favorite presentation of the potential issues when the Singularity will come around remains Stanislaw Lem's "The Cyberiad". Or we could continue focusing on technicalities only and call it a day.
- TheOtherHobbes 12y ago
  Intelligent behaviour in humans and other animals shows evidence of wide-scale situational awareness, and independent, unpredictable motivation. Before you use pattern recognition (i.e. pre-defined, remembered, or improvised abstraction for invariants) to solve a problem, you have to understand there's a problem to be solved. ML seems to be doing well at pattern recognition, but that second problem - understanding what needs to be done in a situation, and why, and accurately predicting possible developments and consequences - is possibly more critical for real singularity-style intelligence. It's machine reasoning as opposed to machine learning - in the street-wise sense, not the theorem proving sense.
shock-value 12y ago
Facial recognition is exactly the kind of "simple pattern matching" that I'm talking about. It's definitely useful, and definitely fits the label "artificial intelligence" (at least compared to other solved tasks), but it simply is just one minuscule piece of what actually constitutes human intelligence. Bear in mind that the original post referenced the "singularity" -- when artificial intelligence is smarter and more capable than actual human intelligence. To be clear, I'm not saying that new AI paradigms are needed to improve that facial recognition score from 97.25% to 99+% (the existing techniques might actually be optimal anyway). I'm saying they are needed to even be able to attempt other tasks that actual humans find effortless, such as interpreting literature, creating a song, carrying on a conversation, decorating a room in an appealing way, understanding symbolic math, having this debate with you, etc. (Certainly symbolic math software has been developed, but no artificial intelligence has been developed that "understands" it and can put it into larger context, particularly one based on neural networks a la the real human brain.) I don't think my thinking is outdated (or even that controversial). It's mostly informed by a Deep Learning class I recently took taught by Yann LeCun, who is definitely at the forefront of the field (although these opinions are entirely my own). EDIT: Also, when I mention new "AI paradigms", I do expect them to build on the deep neural networks currently being studied. But more sophisticated architectures than feedforward nets or even the simple recurrent nets that I've seen so far will be needed, I think. And that will likely require some breakthroughs in theory as well as engineering.