[Xmca-l] Re: Interesting article on robots and social learning

Douglas Williams djwdoc@yahoo.com
Mon Jul 16 19:01:16 PDT 2018


 Hi, Michael--I think it could be, as there is certainly an interest in dealing with bias, especially once you move away from the relatively easily detectable ones in chatbots. 
Frankly, I was thinking in part to check in with you guys to see what you thought, as the questions Kate Crawford poses here in the Neural Information Processing Conference keynote last year are precisely the ones of perspective and mind that I associate with CHAT. Perhaps the most useful thing I can do is to put this in front of you all for consideration:
The Trouble with Bias - NIPS 2017 Keynote - Kate Crawford #NIPS2017


| 
| 
| 
|  |  |

 |

 |
| 
|  | 
The Trouble with Bias - NIPS 2017 Keynote - Kate Crawford #NIPS2017

Kate Crawford is a leading researcher, academic and author who has spent the last decade studying the social imp...
 |

 |

 |


Regards,Doug
    On ‎Sunday‎, ‎July‎ ‎15‎, ‎2018‎ ‎05‎:‎26‎:‎23‎ ‎PM‎ ‎PDT, Glassman, Michael <glassman.13@osu.edu> wrote:  
 
 
I wonder if where CHAT might be most interesting in addressing AI are on topics of bias and oppression.  I believe that there is a real danger that AI can be used as a tool for oppression, especially from some of its early uses.  One of the things people discussing the possibilities of AI don’t discuss near enough is that it picks up and integrates biases from the information it receives.  Sometimes this can be interesting such as the program Libratus that beat world class poker players at Texas Hold ‘em.  One of the less discussed aspects is that one of the reasons it was capable of doing this is it picks up on the playing biases of the players it is competing with and integrates them into its decision making process.  This I think is one of the reasons that it has to play only one player at a time to be successful.
 
  
 
The danger is when it integrates these biases into a larger decision making process.  There is an AI program called Northpointe used by the justice department that uses a combination of big data and deep learning to make decisions about whether people convicted of crimes will wind up back in jail.  This should have implications for sentencing.  The program, surprise, tends to be much harsher with Black individuals than white individuals.  Even if you keep ethnicity outside of the equation it has enough other information to create a natural bias.  There are also some of the more advanced translation programs which tend to incorporate the biases of the languages (e.g. mysoginistic) into the translations without those getting the translations realizing it.  AI , especially machine learning, is in many ways a prisoner to the information it receives.  Who decides what information it receives? Much like the intelligence tests of an earlier age people will use AI decision making as being neutral or objective when it actually mirrors back (almost perfectly) those who are feeding it information.
 
  
 
Like I said I don’t see this point raised nearly enough.  Perhaps CHAT is one of the fields in a position to constantly point this out, explore the ways that AI is culturally biases, and those that dominate information flow can easily use it as a tool for oppression.
 
  
 
Michael
 
  
 
From: xmca-l-bounces@mailman.ucsd.edu <xmca-l-bounces@mailman.ucsd.edu>On Behalf Of Greg Thompson
Sent: Sunday, July 15, 2018 12:12 PM
To: eXtended Mind, Culture, Activity <xmca-l@mailman.ucsd.edu>
Subject: [Xmca-l] Re: Interesting article on robots and social learning
 
  
 
And I'm still curious if any others out there might have anything to contribute to Doug's query regarding what CHAT theory (particularly developmental theories) might have to offer thinking about AI?
 
  
 
It seems an interesting question to think through even if you aren't on board with the larger AI project...
 
  
 
-greg
 
  
 
On Sun, Jul 15, 2018 at 10:55 AM, Andy Blunden <andyb@marxists.org> wrote:
 

I think we go back to Martin's earlier ironic comment here, Michael.
 
Andy
 
Andy Blunden
http://www.ethicalpolitics.org/ablunden/index.htm
 
On 15/07/2018 9:44 AM, Glassman, Michael wrote:
 

The Turing test, at least the test he wrote in his article, is actually a big more complicated than this, and especially poignant today.  Turing’s test of whether computers are acting as human was based on an old English game show called The Lying Game (I suppose one of the reasons for the title of the movie on Turing, though of course it had multiple meanings.  But for some reason they never mentioned the origin of the phrase in the movie).  Anyway in the lying game the contestant had to listen to two individuals, one of whom was telling the truth about the situation and one of whom was lying. The way Turing describes it, it sounds quite brutal.  The contestant had to figure out who the liar was (there was a similar much milder version years later in the US). Anyway Turing’s proposal, if I remember correctly, was that a computer could be considered thinking like a human if the comp the contestant was listening to was lying and he or she couldn’t tell. In essence the computer would successfully lie.  Everybody think Turing believed that computers would eventually think like humans but my reading of the article was that he had no idea, but as the computer stood at the time there was no chance.
 
 
 
The reason this is so poignant is the Mueller indictments that came down yesterday.  For those outside the U.S. or not following the news the indictments were against Russian military leading a scheme to convince individuals of lies about various actor in the 2016 election (also times release of information and breaking in to voting systems).  But it is the propagation of lies by robots and people believing them that interests me.  I feel like we aren’t putting enough thought into that.  Many of the people receiving the information could not tell it was no from humans and believed it even though in many cases it was generated by robots, passing it seems to me Turing’s test.  How and why did this happen? Of course Turing died before the Internet so he couldn’t have known about it.  But I wonder if part of the reason the robots were successful is that they have the ability to mine, collect and aggregate people’s biases and then reflect them back to us.  We tend to engage, believe things in the contexts of our own biases.  They say in salesmanship that the trick is figuring out what people want to here and then couching whatever you want to see in that.  Trump is a master of reading what a group of people want to hear at the moment, their biases, and then mirroring it back to them
 
 
 
If we went back to the Chinese room and the person inside was able to read our biases from our messages would they then be human.  
 
 
 
We live in a strange age.
 
 
 
From:xmca-l-bounces@mailman.ucsd.edu<xmca-l-bounces@mailman.ucsd.edu>On Behalf Of Andy Blunden
Sent: Saturday, July 14, 2018 8:58 AM
To: xmca-l@mailman.ucsd.edu
Subject: [Xmca-l] Re: Interesting article on robots and social learning
 
 
 
I understand that the Turing Test is one which AI people can use to measure the success of their AI - if you can't tell the difference between a computer and a human interaction then the computer has passed the Turing test. I tend to rely on a kind of anti-Turing Test, that is, that if you can tell the difference between the computer and the human interaction, then you have passed the anti-Turing test, that is, you know something about humans.
 
Andy
 
Andy Blunden
http://www.ethicalpolitics.org/ablunden/index.htm
 
On 14/07/2018 1:12 PM, Douglas Williams wrote:
 

Hi--
 
I think I'll come out of lurking for this one. Actually, what you're talking about with this pain algorithm system sounds like a modeling system that someone might need to develop what Alan Turing described as a P-type computing device. A P-type computer would receive its programming from inputs of pleasure and pain. It was probably derived from reading some of the behavioralist models of mind at the time. Turing thought that he was probably pretty close to being able to develop such a computing device, which, because its input was similar, could model human thought. The Eliza Rogersian analysis computer program was another early idea in which the goal was to model the patterns of human interaction, and gradually approach closer to human thought and interaction that way. And by the 2000's, the idea of the "singularity" was afloat, in which one could model human minds so well as to enable a human to be uploaded into a computer, and live forever as software (Kurzweil, 2005). But given that we barely had a sufficient model of mind to say Boo with at the time (what is consciousness? where does intention come from? What is the balance of nature/nurture in motivation? Speech utterances? and so on), and you're right, AI doesn't have much of a theory of emotion, either--the goal of computer software modeling human thought seemed very far away to me.
 
 
 
At someone's request, I wrote a rather whimsical paper called "What is Artificial Intelligence?" back in 2006 about such things. My argument was that statistical modeling of human interaction and capturing thought was not too easy after all, precisely because of the parts of mind we don't think of, and the social interactions that, at the time, were not a primary focus. I mused about that in the context of my trying to write a computer program by applying Chomsky's syntactic structures to interpret intention of a few simple questions--without, alas, in my case, a corpus-supported Markov chain logic to do it. Generative grammar would take care of it, right? Wrong.
 

So as someone who had done a little primitive, incompetent attempt at speech modeling myself, and in the light of my later-acquired knowledge of CHAT, Burke, Bakhtin, Mead, and various other people in different fields, and of the tendency of people to interact through the world through cognitive biases, complexes, and embodied perceptions that were not readily available to artificial systems, I didn't think the singularity was so near.

The terrible thing about computer programs is that they do just what you tell them to do, and no more. They have no drive to improve, except as programmed. When they do improve, their creativity is limited. And the approach now still substantially is pattern-recognition based. The current paradigm is something called Convolutional Neural Network Long Short-Term Memory Networks (CNN/LSTM) for speech recognition, in which the convolutional neural networks reduce the variants of speech input into manageable patterns, and temporal processing (temporal patterns of the real wold phenomena to which the AI system is responding). But while such systems combined with natural language processing can increasingly mimic human response, and "learn" on their own, and while they are approaching the "weak" form of artificial general intelligence (AGI), the intelligence needed for a machine to perform any intellectual task that a human being can, they are an awfully long way from "strong" AGI--that is, something approaching human consciousness. I think that's because they are a long way from capturing the kind of social embeddedness of almost all animal behavior, and the sense in which human cognition is embedded in the messy things, like emotion. A computer algorithm can recognize the patterns of emotion, but that's it. An AGI system that can experience emotions, or have motivation, is quite another thing entirely.

I can tell you that AI confidence is still there. In raising questions about cultural and physical embodiment in artficial intelligence interations with someone in the field recently, he dismissed the idea as being that relevant. His thought was that "what I find essential is that we acknowledge that there's no obvious evidence  supporting that the current paradigm of CNN/LSTM under various reinforcement algorithms isn't enough for A AGI and in particular for broad animal-like intelligence like that of ravens and dogs."

But ravens and dogs are embedded in social interaction, in intentionality, in consciousness--qualitatively different than ours, maybe, but there. Dogs don't do what you ask them to, always. When they do things, they do them for their own intentionality, which may be to please you, or may be to do something you never asked the dog to do, which is either inherent in its nature, or an expression of social interactions with you or others, many of which you and they may not be consciously aware of. The deep structure of metaphor, the spatiotemporal relations of language that Langacker describes as being necessary for construal, the worlds of narrativized experience, are mostly outside of the reckoning, so far as I know (though I'm not an expert--I could be at least partly wrong) of the current CNN/LSTM paradigm. 

My old interlocutor in thinking about my language program, Noam Chomsky, has been a pretty sharp critic of the pattern recognition approach to artificial intelligence.

Here's Chomsky's take on the idea:

http://languagelog.ldc.upenn.edu/myl/PinkerChomskyMIT.html

And here's Peter Norvig's response; he's a director of research at Google, where Kurzweil is, and where, I assume, they are as close to the strong version of artificial general intelligence as anyone out there...

http://norvig.com/chomsky.html

Frankly, I would be quite interested in what you think of these things. I'm merely an Isaiah Berlin fox, chasing to and fro at all the pretty ideas out there. But you, many of you, are, I suspect, the untapped hedgehogs whose ideas on these things would see more readily what I dimly grasp must be required, not just for achieving a strong AGI, but for achieving something that we would see as an ethical, reasonable artificial mind that expands human experience, rather than becomes a prison that reduces human interactions to its own level. 

My own thinking is that lately, Cognitive Metaphor Theory (CMT), which I knew more of in its earlier (now "standard model') days, is getting even more interesting than it was. I'd done a transfer term to UC Berkeley to study with George Lakoff, but we didn't hit it off well, perhaps I kept asking him questions about social embeddedness, and similarities to Vygotsky's theory of complex thought, and was too expressive about my interest in linking out from his approach than folding in. It seems that the idea I was rather woolily suggesting to Lakoff back then has caught on: namely, that utterances could be explored for cultural variation and historical embeddedness, a form ofsocial context to the narratives and metaphors and blended spaces that underlay speech utterances and thought; that there was a degree of social embodiment as well as physiological embodiment through which language operated. I thought then, and it looks like some other people now, are thinking that someone seeking to understand utterances (as a strong AGI system would need to do) really, would need to engage in internalizing and ventriloqusing a form of Geertz's thick description of interactions. In such forms, words do not mean what they say, and can have different affect that is a bit more complex than I think temporal processing currently addresses. 

I think these are the kind of things that artificial intelligence would need truly to advance, and that Bakhtin and Vygotsky and Leont'ev and in the visual world, Eisenstein were addressing all along...
 
And, of course, you guys.
 
 
 
Regards,
 
Douglas Willams
 
 
 
 
 
 
 
On Tuesday, July 3, 2018, 10:35:45 AM PDT, David H Kirshner<dkirsh@lsu.edu> wrote: 
 
 
 
 
 
The other side of the coin is that ineffable human experience is becoming more effable.
 
Computers can now look at a human brain scan and determine the degree of subjectively experienced pain:
 
 
 
In 2013, Tor Wager, a neuroscientist at the University of Colorado, Boulder, took the logical next step by creating an algorithm that could recognize pain’s distinctive patterns; today, it can pick out brains in pain with more than ninety-five-per-cent accuracy. When the algorithm is asked to sort activation maps by apparent intensity, its ranking matches participants’ subjective pain ratings. By analyzing neural activity, it can tell not just whether someone is in pain but also how intense the experience is.
 
 
 
So, perhaps the computer can’t “feel our pain,” but it can sure “sense our pain!”
 
 
 
Here’s the full article:
 
https://www.newyorker.com/magazine/2018/07/02/the-neuroscience-of-pain
 
 
 
David
 
 
 
From:xmca-l-bounces@mailman.ucsd.edu<xmca-l-bounces@mailman.ucsd.edu>On Behalf Of Glassman, Michael
Sent: Tuesday, July 3, 2018 8:16 AM
To: eXtended Mind, Culture, Activity <xmca-l@mailman.ucsd.edu>
Subject: [Xmca-l] Re: Interesting article on robots and social learning
 
 
 
 
 
 
 
It seems like we are still having the same argument as when robots first came on the scene.  In response to John McCarthy, who was claiming that eventually robots can have belief systems and motivations similar to humans through AI John Searle wrote the Chinese room.  There have been a lot of responses to the Chinese room over the years and a number of digital philosopher claim it is no longer salient, but I don’t think anybody has ever effectively answered his central question.
 
 
 
Just a quick recap.  You come to a closed door and know there is a person on the other side. To communicate you decide the teacher the person on the other side Chinese. You do this by continuously exchanging rules systems under the door.  After a while you are able to have a conversation with the individual in perfect Chinese. But does that person actually know Chinese just from the rule systems.  I think Searle’s major point is are you really learning if you don’t know why you’re learning, or are you just repeating. Learning is embedded in the human condition and the reason it works so well and is adaptable is because we understand it when we use what we learn in the world in response to others.  To put it in response to the post, does a bomb defusion robot really learn how to defuse a bomb if it does not know why it is doing it.  It might cut the right wires at the right time but it doesn’t understand why and therefore is not doing the task just a series of steps it has been able to absorb.  Is that the opposite of human learning?
 
 
 
What the researcher did really isn’t that special at this point.  Well I definitely couldn’t do it and it is amazing, but it is in essence a miniature version of Libratus (which beat experts at Texas Hold em) and Alphago (which beat the second best Go player in the world).  My guess it is the same use of deep learning in which the program integrates new information into what it is already capable of.  If machines can learn from interacting with other humans then they can learn from interacting with other machines.  It is the same principle (though much, much simpler in this case).  The question is what does it mean.  As we defining learning down because of the zeitgeist.  Greg started his post saying a socio-cultural theorist be interested in this research.  I wonder if they might more likely to be the ones putting on the brakes, asking questions about it.
 
 
 
Michael
 
 
 
From:xmca-l-bounces@mailman.ucsd.edu <xmca-l-bounces@mailman.ucsd.edu>On Behalf Of Andy Blunden
Sent: Tuesday, July 03, 2018 7:04 AM
To: xmca-l@mailman.ucsd.edu
Subject: [Xmca-l] Re: Interesting article on robots and social learning
 
 
 
Does a robot have "motivation"?
 
andy
 
Andy Blunden
http://www.ethicalpolitics.org/ablunden/index.htm
 
On 3/07/2018 5:28 PM, Rod Parker-Rees wrote:
 

Hi Greg,
 
 
 
What is most interesting to me about the understanding of learning which informs most AI projects is that it seems to assume that affect is irrelevant. The role of caring, liking, worrying etc. in social learning seems to be almost universally overlooked because information is seen as something that can be ‘got’ and ‘given’ more than something that is distributed in relationships.
 
 
 
Does anyone know about any AI projects which consider how machines might feel about what they learn?
 
 
 
All the best,
 

Rod
 
 
 
From:xmca-l-bounces@mailman.ucsd.edu<xmca-l-bounces@mailman.ucsd.edu>On Behalf Of Greg Thompson
Sent: 03 July 2018 02:50
To: eXtended Mind, Culture, Activity <xmca-l@mailman.ucsd.edu>
Subject: [Xmca-l] Interesting article on robots and social learning
 
 
 
I’m ambivalent about this project but I suspect that some young CHAT scholar out there could have a lot to contribute to a project like this one:
 
https://www.sapiens.org/column/machinations/artificial-intelligence-culture/
 
 
 
-Greg 
 
--
 
Gregory A. Thompson, Ph.D.
 
Assistant Professor
 
Department of Anthropology
 
880 Spencer W. Kimball Tower
 
Brigham Young University
 
Provo, UT 84602
 
WEBSITE:greg.a.thompson.byu.edu 
http://byu.academia.edu/GregoryThompson
 


This email and any files with it are confidential and intended solely for the use of the recipient to whom it is addressed. If you are not the intended recipient then copying, distribution or other use of the information contained is strictly prohibited and you should not rely on it. If you have received this email in error please let the sender know immediately and delete it from your system(s). Internet emails are not necessarily secure. While we take every care, University of Plymouth accepts no responsibility for viruses and it is your responsibility to scan emails and their attachments. University of Plymouth does not accept responsibility for any changes made after it was sent. Nothing in this email or its attachments constitutes an order for goods or services unless accompanied by an official order form. 
 

 
 

 
 

  
 




 
  
 
--
 
Gregory A. Thompson, Ph.D.
 
Assistant Professor
 
Department of Anthropology
 
880 Spencer W. Kimball Tower
 
Brigham Young University
 
Provo, UT 84602
 
WEBSITE:greg.a.thompson.byu.edu 
http://byu.academia.edu/GregoryThompson
   
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.ucsd.edu/pipermail/xmca-l/attachments/20180717/27c9cb64/attachment.html 


More information about the xmca-l mailing list