Slashdot Mirror
A Peek Inside DARPA's Current Projects
dthomas731 writes to tell us that Computerworld has a brief article on some of DARPA's current projects. From the article: "Later in the program, Holland says, PAL will be able to 'automatically watch a conversation between two people and, using natural-language processing, figure out what are the tasks they agreed upon.' At that point, perhaps DARPA's PAL could be renamed HAL, for Hearing Assistant That Learns. The original HAL, in the film 2001: A Space Odyssey, tells the astronauts how it knows they're plotting to disconnect it: 'Dave, although you took thorough precautions in the pod against my hearing you, I could see your lips move.'"
The Real issue is some idiot programming a computer to defend against someone just trying to turn it off.A computer should be programmed to know it can make mistakes.
To Hell with the Queen of England!
Here's a LOT of stuff to look through....don't tell anyone ;-)
Top Secret Stuff at DARPA. [DARPA]
The DoD funds a huge percentage of AI research, but at the end of the day they're interested in things that can be easily weaponized or used for simple intelligence sifting heuristics. The most fundamentally interesting research in AI is in the humanoid robotics projects such as those at the MIT shop, and it is from these more humanly-modeled projects that anything like HAL could ever issue. Search-digest heuristics like PAL aren't much like humans and will never lead to anything approching a human's contextually rich understanding of the world at large any more than really advanced racecar design will lead to interstellar craft.
The difference, as Searle would say, between Strong (humanlike) AI and Weak (software widget like) AI is a difference of type, not scale.
I am the one true god. However, as an atheist, I don't believe in myself. I guess I have a self-esteem problem.
HAL was programmed to eliminate any possibile failure points in the mission that he could. Through the spaceflight, HAL observed that the humans in the mission were failable (one of them made a suboptimal chess move, a handful of other mistakes were made). HAL had the ability to complete the mission on it's own. Therefore, HAL made the decision, in line with it's programming, to eliminate the human element.
It makes sense, really, when you think about it. And truly, if Dave had just gone along with it and died, HAL would have finished the job perfectly fine.
DARPA: We don't make the things that kill people. We make the things that kill people better. DARPA: We bring good things to life... that are then used to kill people. DARPA: Who do you want to kill today?
I'm in the hole of the broadband donut.
"watch a conversation between two people and, using natural-language processing, figure out what are the tasks they agreed upon."
Anyone care to guess what they plan to use that little gadget for?
There are many tongues to talk, and but few heads to think. -Victor Hugo
http://en.wikipedia.org/wiki/Technological_singula rity
DARPA has yet to acknowledge the project that I was working on 3 years from now in 2010. Last week, January 14, 2012 we will successfully tested the Time Redaction Project. So, I gave myself the plans tomorrow so that I will be submitting them a few years ago to get the grant money. DOD has used this to send a nuke to kill the dinosaurs. I hope it works.
Politics is the art of looking for trouble, finding it everywhere, diagnosing it incorrectly and applying the wrong fix.
[P]erhaps DARPA's PAL could be renamed HAL, for Hearing Assistant That Learns.
Perhaps, but that's not what the orignal HAL stood for. HAL was short for Hueristic ALgorithmal. Arthur C. Clark had to put that into one of his books in the series (2010 IIRC) because lots of people thought he had derived it by doing a rot25 on IBM.
0*0
00*
***
It was some years ago, but I think that the book also stresses the problem for HAL in that the full mission was never revealed to the human crew, which meant that even too good thinking on their part, at the wrong point in time, would be considered a failure. HAL was programmed/ordered to obey the crew, but also respect the mission objectives, and the contradiction only grew worse.
The only programming that leads to context-rich understanding that could be called "knowledge" in the human sense is self-programming. Like babies. We're all born with a some basic software and a lot of hardware, but it's interaction over time with our environment that we self-program. One might call it learning, but it's more fundamental than just accumulating facts: it's self-creation.
Dennett calls us self-created selves. Any AI more than superficially like a human would be the same.
I am the one true god. However, as an atheist, I don't believe in myself. I guess I have a self-esteem problem.
Later in the program, Holland says, PAL will be able to "automatically watch a conversation between two people and, using natural-language processing, figure out what are the tasks they agreed upon."
PAL's role here is not clear. The 'easier' task would be to monitor the body language of the two conversers and, by lining up a list of tasks with the observation of their head movements, correctly predict which points in the conversation were the ones where someone performed an "agreement" gesture.
The much, much more difficult task would be to actually read lips. There are only certain properties of phonemes you can deduce from how the lips and jaw move; many, many other features of speech are lost. Only when you supply the machine with a limited set of words in a limited topic domain do you get good performance; otherwise, you're grasping at straws. And then taking out most of the speech signal? Please.
But no, DARPA is cool and will save all the translators in Iraq (by 2009, well before the war ends.) PR and vaporware win the day!
Intelligent Design: because MATH is HARD.
> Later in the program, Holland says, PAL will be able to "automatically watch a conversation between two people and, using natural-language processing, figure out what are the tasks they agreed upon."
Sure. You will first have to solve the problem of understanding natural language in an open ended environment - something that computers are not very good at yet.
Quite frankly, AI people have been promising this kind of stuff for some 40 years now, and they have so far been unable to deliver. When is PAL going to be able to do what Holland aspires to? Not any time soon - most likely not within the next 20 years.
AI people, please stop announcing such pie in the sky projects.
WRT to "watch a conversation between two people and, using natural-language processing, figure out what are the tasks they agreed upon":
Here's a link to the actual research that they are likely talking about:
http://godel.stanford.edu/~niekrasz/papers/PurverE hlenEtAl06_Shallow.pdf
As you might expect, the ComputerWorld article's summary of the technology is rather optimistic. Nonetheless, this stuff really does exist, and shows some promise in both military and general applications.
I don't mean to imply humanoid robotics qua robotics is necessary to AI development. Rather, only in a creature that acts as an agent inhabiting the world at large can one expect anything like human-level understanding thereof to develop. It's all very well to develop clever as-if software widgets to simulate understanding in carefully controlled circumstances, but they won't scale to true global context richness because 1) they interact with the world over narrow modalities and 2) they don't have the rich internal structure necessary on which predicate agents with deep and flexible competencies.
It's like we build ever more elaborate visual perception analogues, but they backend into databases that only ask for enumerations of objects discriminated. I don't care how competent the visual system is, it's never going to achieve sentience because it's not part of a whole agent that travels around (in some sense), processes the answers it's getting from the visual system in a multimodal way related to the agent's goals, edits those goals based on new information and so on. It can't just see, it has to look, and it can't just look because someone typed in a domain name, it has to look for a reason and the reason has to be a reason in the sense of being the result of a decision or discrimination, not just an action with a physical cause.
It would seem that the easiest way to allow for all that is to build something that really moves around in the real world. In short, building a robot with all the appropriate competencies might be really hard, but it's still the most tractable way to achieve Strong AI.
I am the one true god. However, as an atheist, I don't believe in myself. I guess I have a self-esteem problem.
This would only work for conversations between people of the same sex. There has never been a conversation between a man and a woman in which both participants would agree on the tasks...
M: Want to continue this conversation at my place?
F: Take a leap!
Computer: Agreed to move conversation to male's residence by leaping.
F: When are you going to mow the lawn?
M: Yeah, I'll get right on that.
Computer: Male agreed to mow lawn at earliest opportunity
I didn't mean to imply that only Strong AI is militarily useful. In fact, I would say that Strong AI is *not* useful, if one thinks about the ethics of forcing anything sentient to go to war in one's place.
Also, I have no trouble recognizing that cleverly-designed "Weak" AI is nonetheless quite strong enough in more conventional senses to be a monumental aid to human problem solving, in the same manner and to the same degree as an ICBM is a great aid to human offensive capabilities.
I am the one true god. However, as an atheist, I don't believe in myself. I guess I have a self-esteem problem.
Except that HAL was paranoid. That the astronauts had a conversation they tried to hide from HAL was more than enough. The actual content of the conversation was immaterial.
As I interpreted the scene: Though the audio pickups were off, HAL had a clear view. So he zoomed in on their faces, panned back-and-forth between speakers, and got a clear shot of their faces - lips, eyes, eyebrows, and other facial markings - as each spoke.
Which tells me he was lip-reading. (Also face-reading.) He knew every word they said and had the bulk of the visual side-channel emphasis as well.
If all he needed to know was that they WERE having a conversation, he could have gotten that from his view through the window, without the camera gyrations.
We, as the audience, got an alternation of the omniscient viewpoint - watching and hearing the conversation - with HAL's viewpoint - silence (except for camera pan and zoom actuators) and an alternating closeups of the two talking heads. Thus we could both follow what was said and observe that HAL was doing so as well - but visually - and was putting a lot of processing power and camera actuator activity into subverting the humans' attemp to keep him from knowing what was said.
Bantam Dominique roosters crow a four-note song. Once you've heard it as "Happy BIRTHday" you can't NOT hear it that way
Yeah but thats if you go by the book.
The book and the movie are two different animals. The movie made no mention of the "Hal was following orders" subplot. Short of saying it outright, the movie makes it pretty clear that Hal screwed up and his pride demanded that he eliminate the witneses. Which, if you ask me, makes a more interesting story.
After reading all the books, I came to the conclusion that Clarke would have been better served by sticking to the screenplay.