More Docs, More Data
A fascinating discussion has taken hold among Robot readers, writers, and editors. Beginning in the July/August issue, continuing with emails and the article “Intelligent Machines Plausible or Fantasy?” in the November/December issue, a significant and growing train of thought emerges. Where does software end and biology begin? What is the difference between a grown machine and a natural one? Can Mary Shelleys character, Dr. Frankenstein, ever emerge in reality by creating artificial life?
Robot continues pursuing this topic by interviewing two prominent authorities in the field of computer science and human thought processes. Dr. Michael Mauldin is a widely recognized authority on both information retrieval and artificial intelligence. He holds five patents, including those for the Lycos search engine and a virtual personality. He has researched and developed on-line entities as part of his research on the Turing Test (in which judges remotely converse with humans and machines, attempting to identify which represent artificial intelligence through semantic content and speech patterns.)
Dr. Christopher Hale is a Senior Research Scientist at the Georgia Tech Research Institute (GTRI). Dr. Hale serves as the lead for Human-systems Integration (HSI) and cognitive systems engineering in GTRIs Dayton, Ohio office. He has held research and systems engineering positions in academic, industry and defense settings. For the past 30 years he has been involved in diagnostic problem solving and explanationbased learning; human-systems integration and performance/decision support systems; and design/ evaluation of human-computer interaction, controls and displays.
ROBOT: First off, lets tackle the broadest question. What would a /images/drupalliving machine be like?
DR. MAULDIN: Asimovs definition of a robot is a machine that can sense its surroundings, make decisions, and in turn affect its surroundings. Perhaps the simplest definition of living is something that can interact with its environment, grow, adapt, and reproduce. A living robot therefore would have all of these characteristics.
DR. HALE: Well, responding strictly to the term /images/drupalliving, without any consideration of consciousness or self-awareness; I would say a living machine would have to satisfy six criteria outlined by Kemeny: theyd have to possess some kind of metabolism, exhibit locomotion and reproducibility, show some degree of individuality, exhibit intelligence, and arise from natural composition. Now, defining all these criteria is problematic in themselves, so I wont try that here. One of the complications of wrestling with the /images/drupalliving machine issue is that the answer depends on your frame of reference. Biologists will answer differently than physicists, for example. And then philosophy would probably throw in a bunch of stuff about volition and intentionality.
DR. MAULDIN: Well clearly the /images/drupalnatural composition requirement is simply stacking the deck against alternative forms of existence. And demanding a metabolism is also focusing on /images/drupalhow something lives rather than addressing the core question of being alive.
To me, the question of whether a machine is /images/drupalalive is far less interesting that whether that machine is /images/drupalintelligent. There is a constant theme in science fiction (our modern mythology) about intelligent machines. One important milestone is when the machines start to pursue their own goals, rather than those of their human creators.
DR. HALE: I guess the only thing I would add here is a criterion that physicists often talk about: That living things, at least temporarily, resist the second law of thermodynamics/images/drupalthat they resist the tendency toward entropy. Maybe just another way of saying that staying alive means certain kinds of interactions with ones environment. What kinds? Seems now were sliding over into self-awareness and teleology. Are living machines, then, necessarily self-aware?
ROBOT: The technology almost certainly exists today to connect enough processors, memory, fuzzy logic algorithms, and databases together that its not too far fetched that we could simulate the mental functions of higher level organisms. How could someone know if they were talking via keyboard or cell phone to a simulated intelligence vs. a human being?
DR. MAULDIN: The Turing Test, first proposed by Alan Turing in 1950 as a thought experiment, has been run several times, perhaps most famously by the Loebner Prize Foundation. In 1991, for example, my Chatterbot program Julia was judged /images/drupalmore human than one of the actual human confederates by one of the judges. All that means is that it is easy to fool some of the people some of the time. To date no known program can fool most of the people most of the time.
While the current ability of conversational programs is more than adequate to fool the unsuspecting questioner, a knowledgeable person can easily determine a computer program by asking questions that require integration of multiple facts including real world knowledge and statements made by the questioner:
Judge: Paul is a quadriplegic who likes games and sports.
Program: I like games and sports, too.
Judge: Do you think Paul would like to play football?
Thats not the kind of verbal trap youd lay for someone in normal conversation, but its the kind of question that might have been asked by Deckard during a Voight-Kampff test in Blade Runner.
DR. HALE: The Turing test, eh? When you call your mother on Sunday nights, how do you know youre talking to her and not someone or something else? Words like context and appropriateness and generative and constructive are coming to mind. Note that these go beyond the mere parlor trick of a computer fooling us with its /images/drupalintelligence. Weizenbaum (creator of a conversational computer program) already did that. Its not so much what gets output as the way the output is generated. My view of intelligence is that there is a /images/drupalconnection to a context, paired with an appropriate generative algorithm that leads to truly intelligent human behavior. Lets turn the test around a little: What if the entity on the other end of the line responds to a comment about your vocation with:
/images/drupalShall we rapidly or leisurely talk a little? In spite of your disgusting interest in art, listen to this. The days have dulled since the apex of my dissertation. The commitment was ill-timed and an internist should not have been involved. A captivating tale was told, but then it slinked slowly away. This door is only for you, and only you may enter.
So, who are you talking to a computer? A psychiatric patient? Franz Kafka? More importantly, how do you know?
DR. MAULDIN: The Turing Test represents an achievement of intelligence because we humans pride ourselves on language to distinguish ourselves from other inhabitants of the planet. We push the boundaries by studying the language abilities of birds, dolphins, and primates to reassure ourselves that our language makes us human and unique.
But consider the recurring theme of an intelligent computer program like Colussus, Proteus or SkyNet. Theres no reason to expect the language of a truly intelligent computer to be decipherable by humans.
As another example, once a neural network has been trained to perform a given task, there is no way to look inside and see just how the computer solves the problem. The 1s and 0s are just as patternless to us as the grey matter of a brain being dissected.
DR. HALE: It seems to me that there is some kind of presence and participation, a common /images/drupalfield, that both parties have to inhabit for a simulated intelligence to genuinely fool an astute observer. Something emerges from a mutually participative, intelligent interaction that signals to both parties that such an interaction is taking place that both parties actually understand what is taking place.
ROBOT: What if we take the previous question, and dumb the being down to a simple organism like a fish, lizard, or even a housefly? Can simulating enough behaviors, outside pure reproduction, qualify a machine as /images/drupalliving?
DR. MAULDIN: To my mind, simply reproducing in a natural environment qualifies a being as alive. Any additional skills are just a bonus.
DR. HALE: Are there any counterexamples, something that is alive solely on the basis of reproduction that does not possess any of the other attributes of /images/drupalaliveness? If not, then maybe reproduction is the only criterion we need.
DR. MAULDIN: There are two thresholds of complexity; at one level of behavior we describe the organism as /images/drupalalive. At some much higher level of complexity we use the word /images/drupalintelligent. When people use the word /images/drupalintelligent, they really mean /images/drupalmysterious. In the mid twentieth century chess was a pastime viewed as highly intellectual, and chess masters were considered very intelligent. Naturally programmers set to work to create programs that could play chess well enough to be considered intelligent.
But once those programs were good enough to beat most humans, a quick look under the hood revealed that the computer was using easily understood methods such as enumeration and trial-and-error. Since the mechanisms were not mysterious, the task of playing chess lost its appeal as an example of intelligence. Indeed, once any computer is programmed to perform a given, that task is soon demoted to being /images/drupalmerely mechanical. The computer never gets the reward of being labeled /images/drupalintelligent
When we consider certain animals intelligent, it really means that their behavior is complex enough to surprise us. Insects are predictable and therefore not intelligent. But above a certain threshold, we use word intelligent. Mammals get lots of credit for being intelligent, reptiles not so much.
It should be obvious that there are two characteristics needed for machine intelligence: first the machine must be complex enough to solve problems in ways that surprise the observer, and secondly the mechanisms used must be carefully hidden from view to avoid the /images/drupalmoving goal line problem.
DR. HALE: I go back to Kemenys criteria. In principle, it seems like we should be able to do it if we can attain some kind of volitional or teleological behavior, paired with the tight connection to the environment that I mentioned earlier. By /images/drupaltight connection to the environment, I mean in the dynamical systems sense. Humans typically are exquisitely tuned to the structure and constraints of their environment and thats a large part of being alive, conscious and self-aware. In fact, when these connections are broken, we often define that as psychosis and organic dementia.
ROBOT: In the 1999 movie Bicentennial Man, a mechanical, intelligent robot gradually converted himself into an organic being. He was judged to be human when he converted the last mechanical part into organics, enabling him to age and die. In real life, we are gradually replacing biological parts in humans with mechanical ones (joints, bones, tendons) and even robotic prosthetics. Where do you think a being becomes, or ends, being a /images/drupalhuman?
DR. MAULDIN: For me being human is all about the mind, but that mind is shaped by our physical needs. Our need to reproduce entails our being social with other humans. A human being with an artificial limb is no less human, but a brain in a jar with a digital interface seems too disconnected from society to be human.
DR. HALE: I understand the sentiment here, but Im uncomfortable with the connectedness criterion. Psychotics are not well-connected to society, nor are Alzheimers patients. Some socially adept people dont reproduce. Mind certainly has a physical basis, but while there is certainly a biological aspect to this, but lets not forget the rest of the puzzle. Emotions and culture might be just as important as biology. There was a time when Native Americans and African Americans were considered sub-human. There was the holocaust in World War II. At least some people base their support for capital punishment on the implicit notion that the person being executed has committed an inhuman act. If I replace literally every part of me with silicon and titanium (an impossibility) maybe Im no longer human. [Of course, the mechanical parts will wear out, too, so that criterion might not be so airtight]. I think Id be more concerned about the emotion and culture.
DR. MAULDIN: To quote C.S. Lewis: /images/drupalYou dont have a soul. You are a Soul. You have a body. For me, that soul is the software running on your brain, which is just hardware. Replacing the body with sufficient fidelity to keep the program running doesnt change your humanity. But so much of what our software does is tied to our bodys needs, sensations, abilities and flaws that having that body during the development of your soul software is crucial. As a consequence, I dont think that the Bicentennial Man would ever be able to achieve humanity; intelligence perhaps, but humanity never.
ROBOT: If you had a 10 billion dollar annual budget and a critical need to create a useful artificial intelligence with human level performance and flexibility (ie: /images/drupalartificial human) in a decade, give our readers an overview of your technical approach. Biological? Software? Cyborg? Something else?
DR. MAULDIN: Thats a very vague question. There are lots of /images/drupaluseful artificial intelligences in existence already that cost far less than a billion dollars. The DARPA autonomous vehicle challenges spring to mind immediately.
But supposing that I had a specific requirement to create a computer program to perform a given task/images/drupalId use that same competition strategy to motivate multiple teams of researchers. We humans are at our best when we are trying to outdo other humans.
Id set up an annual competition with a gradually increasing performance requirements and gradually increasing prize payoffs.
DR. HALE: This seems like the right incentive structure to try an achieve this goal. In fact, the /images/drupalmarket economy we have in the United States seems to have done pretty well with innovation of the kind needed for this problem as long as one has a decent enough science and engineering infrastructure (read /images/drupaleducational system here) to respond usefully to the incentives.
I would try to pursue three lines of development. From the bottom- up, I would try to begin building up a system based on the hardware of the human so as to properly constrain behavior and performance. From the top down, I would try to find high-level algorithms that provide solution strategies to wide ranges of problems that the entity will encounter in the world. And, finally, these two approaches need to be brought together within a context of adaptively interacting with a dynamic environment.
DR. MAULDIN: Id stick with software, because we already know how to grow biological organisms that achieve human level performance within decades, and at least the initial process is rather enjoyable.
The most important thing about software is that once you have it working you can run additional copies of the software on other machines, multiplying your productivity in an obvious way. If I create an intelligent machine with a biological component, how can I be sure that a copy will do as well as the original?
Dr. Mauldin, http://lazytoad.com/lti/
Words by Kevin Berry