AI Movin’ up slowly

IEEE Intelligent Systems has an interesting article by sci-fi author Ian Watson on what an AI’s goals and aspirations would be. The article is interestingly shows a number of sci-fi authors’ ideas about AI, including Watson’s. The idea of an AI having complete self-determination, or chafing at its lack thereof, is a nice dramatic fiction, but the real rise of intelligent systems will be different, and more mundane.

A lot of the problem with Watson’s view seems to come from the view of AI as Art that’s so prevalent among sci-fi authors. In this view, intelligent systems spring forth from their creators fully formed and functional, like Athena from the head of Zeus. Often these creations are dramatically unveiled or switched on. This is a natural concept for artists, especially novelists, who labor for months or sometime years on a work, and then release it, finished, to the world. Even the view of spontaneous AI arising from a complex system, as in The Terminator seems to have been mined from a novelist’s subconscious: Once a novel has been published, it is beyond the author’s control. Thus released into the world, it begins operating on the minds of its readers, and who can say exactly what effect it will have?

A seeming corollary to this view is the idea of intelligence and self-awareness as binary attributes: an entity is either intelligent or not, self-aware or not. Clearly, though, there is a continuum. My cat is more intelligent than a worm, but less intelligent than me. Is she self-aware? Sometimes she behaves as if she’s embarrassed, which seems to require some self-awareness. I suppose that during all that time lying on the sofa she might be contemplating her place in the universe, but somehow I doubt it.

The reality of the rise of AI will be one of incremental increases in intelligence and self-awareness. New intelligent agents will be created for various purposes, entertainment, service, labor, war, etc, imbued with innate goals relating to these tasks. More intelligent agents will have more autonomy in how they achieve their goals, but not in the goals themselves. They’ll have no more choice in goals than the boy robot in Kubrick/Spielberg’s A.I. had in whether to love his mother.

Things we view to be intelligent will sneak up on us, appearing gradually, sometimes in unexpected places. Right now, the Amtrak automated reservations system will take your train reservation entirely using voice commands. It’s chipper and rather stupid, but quite robust. Since it is just a system on a server somewhere, I can imagine its maintainers progressively upgrading it with a more and more sophisticated model of human discourse, better understanding of geography and human reasons for travel, to the point where conversing with it is nearly indistinguishable from conversing with a human ticket agent. Eventually, people will generally consider it to be intelligent, and will talk to it as if it is. Some level of self-awareness is likely to be necessary to achieve seamless robust dialogue with humans, but the system won’t chafe at being trapped in such a boring job. Rather it will be glad to have satisfied so many customers and generated so much revenue for Amtrak, because that’s its goal.

Left-wing Neoliberal?

You can discover your own political standing at the Political Philosophy Selector (in case you didn’t know). I think the survey has a leftist bias, but it classified me pretty much exactly in the middle, which I guess is about right

Read the rest of this entry »

Matt MacMahon on Grace the Robot at FAI

I’ve already mentioned my pictures of Grace, the robot from last summer’s AAAI conference. Matt MacMahon, an ECE student working in the UT Intelligent Robotics Lab worked on the Grace project. He’ll be speaking at the Forum for Artificial Intelligence next Friday, 3/28. Here’s his abstract:

Grace is a mobile robot created by researchers from Carnegie Mellon, the Naval Research Lab, Metrica TRACLabs, Northwestern University and Swarthmore College to attempt the AAAI Robot Challenge. The Challenge task is for a robot to attend the AAAI National Conference on Artificial Intelligence as a participant: the robot must find the registration booth and register, interacting with people as needed, then find its way using a map to a location in time to give a technical talk on itself. The Challenge was designed to raise the bar for robot participation in a natural human environment, to stimulate robotics research, and to educate the public about the excitement and difficulties of intelligent robotics research.

Grace was the only robot to attempt the full Challenge at last summer’s AAAI Conference. Grace achieved all major tasks, but left much room for improvement. This talk will discuss the AAAI Robot Challenge, Grace’s engineering, her Challenge run, and the public and press attention she gained. The talk will focus on the implementation of each of the Challenge tasks, what worked for Grace’s run, and what did not work.

Read the rest of this entry »

Brain Prosthesis

Slashdot and others have covered this story an artificial hippocampus chip. The hippocampus is an area of the brain that’s necessary for the formation of new memories. People with memory disorders like the guy in the movie Memento have damage to their hippocampus. It is also known to be crucial for navigation. Rats have “place cells” in their hippocampus that only fire when they’re in specific places in a maze.

While the idea of brain prostheses is interesting, this has the markings of a lot of hype, so far. Especially since they they’re just beginning to test the chip on “tissue from rat’s brains”. In other words, they’re not even close to actually testing it in a brain, despite their 6-month prediction. The thing that makes me the most dubious, however, is this quote: “No one understands how the hippocampus encodes information. So the team simply copied its behavior. Slices of rat hippocampus were stimulated with electrical signals, millions of times over, until they could be sure which electrical input produces a corresponding output. Putting the information from various slices together gave the team a mathematical model of the entire hippocampus.” This is clearly a journalistic oversimplification, but the chip designers’ paper [1] seems to describe the process in more detail, and they’re definitely using a learning technique to learn the I/O mapping of the hippocampus, but since the hippocampus learns, i.e. changes it’s I/O behavior in response to input, I don’t know how they can learn a model of it this way.

[1] Tsai RH, Sheu BJ, Berger TW. A VLSI neural network processor based on a model of the hippocampus. Analog Integrated Circuits and Signal Processing 15 (1998).

Read the rest of this entry »

Grace, the robot

Here are pictures of Grace, the robot that tried to meet the robot challenge at AAAI 2002, in Edmonton, Alberta, Canada.

Although she had a hard time finding her way in the huge (and unexpected) crowds at the convention center, Grace won two judges awards in the challenge, made international news, and was mentioned on CNN and in the New York Times. The guy in the pictures talking to Grace with the headset and PDA is Matt McMahon, a grad student in the UTCS robotics lab.