Use the term "Artificial Intelligence" around most people, and it conjures images of thinking, emotive machines, often in anthropomorphic form. Film and fiction have portrayed AI so often and in such depth, that the meme of "machine consciousness" has become embedded in people's minds. From 2001's HAL to Star Wars robots to Terminator and the sad little boy in A.I., we've been provided with images and mythic tales of machines making informed conscious decisions and exhibiting emotion.
Reading consumer-level science journals and corporate press releases can lead one to believe that AI is making huge leaps towards self-aware machines. It is as though mere moments separate us from being able to find out why the answer is 42. [i] Recently, Dr. Douglas Lenat, the AI visionary and world-renowned computer scientist leading the Cyc project, gave interviews that suggested that Cyc had achieved consciousness and that they were busy programming hard logic moral rules for it.
Cyc, (pronounced Psych), is a project working on a "commonsense" approach to AI that has been quietly under development for almost 2 decades. When I asked John DeOliveira, the marketing director at Cycorp, to define "commonsense", he said: "If you look at an encyclopedia, you'll see a great deal of knowledge of the world represented in the form of articles. Common sense is exactly not this knowledge. Common sense is the complement of this knowledge. It is the white space behind the words. It is all of the knowledge that the article writer assumed all of his/her readers would already have prior to reading the article -- knowledge that could be put to use in order to understand the article. Cyc is about representing and automating the white space." (I love that answer.)
Large portions of the Cyc knowledge base will be released to the world at large in August 2001, in the form of OpenCyc, which may be the largest open-source collection of inferentially categorized data in the world. [ii] According to the company, it will have a knowledge base of "6,000 concepts: an upper ontology for all of human consensus reality and 60,000 assertions about the 6,000 concepts, interrelating them, constraining them, in effect (partially) defining them." The main, nonpublic Cyc knowledge base has over 1.4 million assertions and took approximately 500 person-years and $50 million dollars to develop.
Cyc is not the only "commonsense" project out there. MIT and Mindpixel both have projects [iii] that are similar, although Cyc appears to be considerably more sophisticated. Dr. Lenat was quoted in the Los Angeles Times [iv] as saying "Cyc has goals, long and short-range. It has an awareness of itself. It doesn't care about things in the same sense that we do, but on the other hand, we ourselves are only a small number of generations away from creatures who operated solely on instinct." In the same article Dr. Lenat went on to say, "No one ever told HAL that killing is worse than lying. But we've told Cyc."
Well, I don't know about you, but I found that statement kind of spooky. Even with a couple dozen philosophers on board, how do you set non-contextual rules for that sort of thing? Whose morals are you going to use? For example, is all killing worse than all lying? What if it's the state doing the killing in an execution? Is it wrong to kill someone who wants to die? Is it wrong to kill to protect yourself? These are the types of questions I wanted to ask Cyc. If they were building a machine with morals, I wanted to understand what sort of context it was using for moral ontology. [v]
I was curious as to whether I could "talk" to the prime knowledge base, Cyc itself, in Natural Language format. Now, when I say talk to it, I didn't expect it to be an Alicebot chat clone. I'm not sure exactly what I was expecting, but I figured that any machine that had been hand-coded by philosophers had to be the closest we've come to a conscious machine, and I wanted to do an interview with it. (Yes, I can hear you laughing...and it's not very nice.)
In my chat and email conversations with John DeOliveira at Cycorp, he was kind enough not to giggle when I stated my desire to interview Cyc. He provided a lot of information and access to DARPA tests of Cyc so I could see how the query process and information addition worked. But he said that it wouldn't be possible for me to "play" with Cyc. A trained skill-set is required to be able to efficiently interface with the Cyc engine, and naturally, they're reluctant to let random people come bang on a keyboard. Mr. Oliveira answered many questions that I had, but he was reluctant to answer questions about Dr. Lenat's assertion of consciousness, the topic in which I was most interested.
So, I thought perhaps I was just asking the wrong questions. After consultation with many experts, [vi] and reviewing many books, white papers and thesis after thesis; I came to the conclusion that the media's representation of AI is akin to seeing hackers portrayed in the media. Which is to say, that the reality and the meme (or belief) are radically different things.
I also discovered that I was unlikely to get anyone to firmly commit to a discussion about "consciousness" in a real-life application. Which is to say, nobody has actually created a self-aware, thinking machine. A machine which thinks all the time, without being given a set task or problem and has a philosophy about what role it fills in the grand scheme of creation. I couldn't find anyone who was willing to say that there was a project that could pass a Turing [vii] test. Yet. Which may be just as well, as the science of what's really going on in AI is even more interesting and varied than the fiction that surrounds it.
The field has an amazing array of people with different educational backgrounds; from computer scientists and theorists to philosophers, linguists, mathematicians, behavioral scientists, neuroscientists and cognitive psychologists, just to name a few. Ongoing, live AI applications span the gamut from expert systems that help diagnose patients to big-brother facial recognition software being used by various police forces. AI is currently used by any number of corporations to sift through reams of accumulated data, it's being used by people working on translation applications to enable the world to overcome language communication barriers, it's used by the military to help plan and simulate attack and defense strategy. It's being used to build cars and run trains. AI is very real, and very active. What it isn't; is human-shaped and moody.
When I started researching this article, I really had no idea what I was getting in to. I generally consider myself smarter than your average lab rat, but the sheer volume of information out there is astounding, much of it written at the doctorate level and assumes a familiarity with the underlying premises. I think it may be fair to say that I hit information overload and realized that it probably wasn't possible, in a short amount of time, to fully explore the various areas, applications and theories. So, I present an overview of the current state of artificial intelligence, and provide links for your edification should you have a deeper interest in a particular topic. Strap yourself in; it's a dense ride from here on out.
The study of AI can be loosely divided into the following segments: expert systems, natural languages, simulation of human sensory capabilities, robotics, neural networks, data mining and machine learning in general.[viii] But, the segments often overlap like a complex Venn diagram.
Expert and Knowledge-Based Systems [ix]
Expert systems are based on the theory that creating propositions and performing logical transformations upon those propositions can model human experience and expertise. Expert systems are comprised of a knowledge base, a set of algorithms, which define how to infer knowledge, and an inference engine. New facts or answers are derived when the knowledge is fed through the inference engine and is processed according to the algorithmic rules.
Such systems are being used in NASA strategy and management, financial institutions like Price Waterhouse, medical diagnosis, manufacturing plants, distribution fulfillment centers, customer support, routing and dispatch centers, network analysis, network security, and in conjunction with image processing software police video surveillance cameras which record your face for storage and retrieval in a national databank of digitized citizens.
Natural Language Processing [x]
Natural language can be defined as the language acquired by humans in the normal socialization process as opposed to an "invented language" such as predicate calculus, which has a set of specified syntax and explicit semantics. AI experts in Natural Language Processing (NLP) are working towards a goal where the user is able to interface with an application in their own words (ergo, natural language), rather than using code or symbolic logic. Considering the complexity of all human languages, and the innate philosophical issues of semantics, creating a natural language application is no easy feat.
NLP is an area where the divide between rationalists and empiricists is pronounced. Rationalists generally hold the belief that significant amounts of knowledge about processing language in the human brain are not derived from experience but are instead the result of genetic inheritance. This philosophy is seen in NLP applications where enormous amounts of "starting" information and assumptions are hand coded in to simulate a human base, similar to the commonsense approach of Cyc.
The empiricists believe that cognitive ability in the brain allows humans to categorize, sort and evaluate stimuli and that rich sensory input is the basis of the ability to process language. Therefore, an empirical approach would be to suggest a general language model and apply statistical and pattern recognition methodology to a large amount of language.
The search engine, www.AskJeeves.com is an implementation of a natural language application, as are customer-service bots and chatterbots. Universities are working on unified corpus translation and document summarization. Voice recognition, voice synthesis, and written computer interfaces are outgrowths of the field. Many projects in AI have a NLP front-end; hence the field is often imperative to the success of other artificial intelligence projects.
Sensory Capabilities and Digital Processing [xi]
The most common application of this branch of AI is in allowing computers to "see". Applications can process images, streaming video or live action and pick out patterns and matches. For example, at the last Super Bowl, when cameras scanned each fan as they entered the arena, the images were run through an algorithm in the expert system to look for facial matches with known criminals. If the image system found a match, the expert system (using a neural network) queried a knowledge base that attached prior offenses to the matched photos, then sent a digital "package" of information consisting of the photos, suspect id, and prior offenses to a human operator for final analysis and action.
In a less Orwellian vein, Scott Acton of the University of Virginia is working on digital imaging for semiconductor inspection, make sure all the chips are there, that the solder is in the right place, etc. It's also used for the quality grading of T-bone steaks and for pizza manufacturing; is this pizza round and does it have enough pepperoni? The Post Office uses it for sorting and my favorite application of digital imaging and sensory capabilities is the use of AI to manage automatic sheep shearing in Australia.
MIT is doing some amazing robotic sensory work with Cog And Cog leads us into our next category.
Robotics [xii]
This is the section of AI that captures the most public interest, possibly because it's the most tangible incarnation of the science. And quite possibly it is the most inclusive of the branches, as it requires a bit from every other AI segment, with a little fuzzy logic thrown in for color. And robots have the common touch. They are something that a hobbyist can build. Even I have a Lego Mindstorm running about the place.
Robots and mechanized self-aware life has long been a staple of science fiction. And fiction about robots usually works under the following premise; machines which are shaped like humans, move like humans, and interact as humans, must have cognitive process and emotions, like humans. Many, like Phillip K. Dick, envisioned the advent of anthropomorphic thinking machines as an apocalyptic pox let loose on an unsuspecting world.
Whereas huge advancements in robotics have taken place in the last few years, when it comes to practical application, robotics has been more of a success for mechanical engineers than it has for AI researchers. A robot, at its base definition, is a mechanized method of performing a set task. There are robot installations in manufacturing plants and industries all over the world, but they don't wander around on two legs and hold conversations about the existential angst caused by lifting auto frames all day. They are often very large pieces of equipment that follow a set of programmed rules in order to accomplish a given task. They don't think about ways to improve the method or the task itself, they simply follow orders.
MIT's Humanoid Robotics Group and the Waseda University's Humanoid Group are working on robots which may change that.
Artificial Neural Networks [xiii]
Artificial Neural Networks use a set of processing elements (or nodes) loosely analogous to neurons in the brain. These nodes are interconnected in a network that can then identify patterns in data as it is exposed to the data. In a sense, the network learns from experience just as people do. This distinguishes neural networks from traditional computing programs, which simply follow instructions in a fixed sequential order.
Most ANNs have some sort of "training" rule whereby the weights of connections are adjusted on the basis of data. In other words, ANNs "learn" from examples and exhibit some capability for generalization by using fuzzy logic. According to Dr. Applebaum, who wrote her doctoral thesis on applied mathematics and fuzzy logic, "Fuzzy logic deals with vague concepts such an 'almost zero', 'a smooth ride', 'almost bald', and so on, in the context of creating systems that do qualitative reasoning with uncertain and 'vague' data."
In practice, ANNs are especially useful for classification and problems which have lots of training data available, but to which hard and fast rules (such as those that might be used in an expert system) cannot easily be applied.
Data Mining [xiv]
Data mining is the discovery and modeling of hidden patterns in large amounts of data. It is usually case-based, in that the parameters can be statistically modeled. Technically, data mining is statistical analysis, but on a complex scale. IBM invented data mining and holds some of the patents. One of the goals of data mining is to allow the user to discover patterns and build models automatically, without knowing exactly what she's looking for.
The models are both descriptive and prospective. They address why things happened and what is likely to happen next. A user can pose "what-if" questions to a data-mining model that can not be queried directly from the database or warehouse. Examples include: "What is the expected lifetime value of every customer account," "Which customers are likely to open a second account," or "Will this customer cancel our service if we introduce higher fees?" (Questions like this assume a Natural Language front end.)
Text mining is a subset of data mining which applies the same rules and logic, but is directed at gleaning information from large bodies of text rather than numerical data. The information technologies associated with making data mining a functional process are neural networks, genetic algorithms, fuzzy logic, and rule induction. Data mining is becoming more prevalent as businesses, governments and organizations look for ways to leverage the existing mountains of information they already have.
Machine Learning in General. [xv]
Machine Learning is the study of computer algorithms that improve automatically through experience. In other words, it's an application that can learn from experience, from observation and from direct input. Neural networks are an example of machine learning, as are some image recognition projects. Current applications include credit card fraud detection based on past spending patterns, targeted marketing campaigns based on past purchase patterns, customer service based on products purchased, crop disease diagnosis, human medical diagnostics and last, but not least, games. Probably the most famous example of a machine which taught itself, IBM's Deep Blue, which taught itself to play chess better than the human world chess champion, Garry Kasparov
Well, if you're still with me, that brings us to the end of the overview. I set out to understand how one project was defining morals for a theoretically "conscious" machine and ended up doing broad look at the sciences and disciplines that are encompassed by the term "artificial intelligence". (Talk about scope-creep.) I never did figure out how morals could be coded into a non-contextual frame, but I did discover that I agree with Julian Jaynes [xvi]; consciousness may be a neurotic illusion.
As you can see, the science of AI is working towards some amazing stuff, and going in fascinating directions. The people who are working on it, must have bigger brains than the rest of us, I'm sure of it. Otherwise they would just run around looking dazed from all the information coming at them. And while I believe the study and application of AI may be one of the cultural pinnacles of achievement, it is by no means anywhere near the mythical standards held up by the popular media.
In fact, I would suggest that the media actually does a disservice to the scientists by ignoring the real science in favor of imaginary, but flashier science. The real science is moving forward, perhaps fraught with ethical implications, perhaps blindly into avenues for which we don't have philosophical signposts yet, but moving forward none the less. The twenty-first century is upon us, and these scientists are leading the charge into the future. The question becomes, can the philosophers and ethicists keep up?
Links, References and More Information
This is by no means an inclusive survey of all technologies or applications in the field of AI. Please see the links below if you are interested in researching more about the field.
AI's Greatest Trends and Controversies. IEEE Intelligent Systems. (January/February 2000). (Also available in pdf.)
AI TopicsEndNotes
This site is oriented toward teachers, students and the general public interested in or researching artificial intelligence or expert systems.American Association for Artificial Intelligence
Professional organization.Artificial Intelligence Resources
A links list from the Institute for Information Technology. Includes bibliographies, publications, societies, newsgroups and mailing lists, subject lists, etc.Journal of Artificial Intelligence Research
Many articles available full-text.MIT Artificial Intelligence Laboratory
Resources and reference section of this web site will be primarily useful to researchers.
[iii] The OpenMind project and the Mindpixel project are creating their projects by accepting input from "normal" people, rather than knowledge experts. The goal is to create a common database of knowledge that is generally accepted as "true" by most people.
[iv] Hiltzik, Michael. "Birth of a Thinking Machine"Los Angeles Times 06/21/2001. Link to article not freely available as the article resides in a pay per view archive at the Times. Archive searches cannot be saved, so I can't give you that link either. However; you can surf to: www.latimes.com/services/site/archives/ and search for either the title or the author and get the article for a nominal fee, or your local library may have an archived copy of the newspaper.
[v] Ontology is a commonly used word in both philosophy and in artificial intelligence, but it has a different definition in each field.
1. philosophy: A systematic account of Existence.For AI systems, what "exists" is that which can be represented. When the knowledge about a domain is represented in a declarative language, the set of objects that can be represented is called the universe of discourse. We can describe the ontology of a program by defining a set of representational terms. Definitions associate the names of entities in the universe of discourse (e.g. classes, relations, functions or other objects) with human-readable text describing what the names mean, and formal axioms that constrain the interpretation and well-formed use of these terms. Formally, an ontology is the statement of a logical theory.2. artificial intelligence: (From philosophy) An explicit formal specification of how to represent the objects, concepts and other entities that are assumed to exist in some area of interest and the relationships that hold among them.
A set of agents that share the same ontology will be able to communicate about a domain of discourse without necessarily operating on a globally shared theory. We say that an agent commits to an ontology if its observable actions are consistent with the definitions in the ontology. The idea of ontological commitment is based on the Knowledge-Level perspective.
3. information science: The hierarchical structuring of knowledge about things by subcategorizing them according to their essential (or at least relevant and/or cognitive) qualities. See subject index. This is an extension of the previous senses of "ontology" (above) which has become common in discussions about the difficulty of maintaining subject indices.Source: The Free On-line Dictionary of Computing, © 1993-2001 Denis Howe
[vi] This article could never have been written without the help of the following experts in the field. I was lost indeed until they stepped in with advice, comments, reference materials and cleanup offers when my head exploded..
Dr. Ellen Applebaum: fields of specialty include applied mathematics and fuzzy modeling.
Dave Evans: Ph.D. candidate at Columbia University. Some NLP projects at Columbia University.
Dr. Elaine Kant: SciComp Inc. http://www.scicomp.com/
Dr. Karin Vespoor: Computational Linguist
Kelley Ann Newton Kranzler: KANconsulting (Kelly was great fun and a great help. She spent hours with me discussing the various branches of AI. I wish I'd had the space to include all of her witticisms and examples.)
Amanda Holland-Minkley: Ph.D. candidate at Cornell University working on natural language generation. Her projects may be viewed here and her home page may be viewed here.
[Viii] Dr. Ellen Applebaum, a respected and well-published expert on AI, suggested a different way of looking at the paradigm of AI. In an email interview, she said: "I actually see two different major paradigms in AI. One is the "rational" school of thought of what AI is and isn't. The other is the "empirical" school of thought - more feeling based. The rational school of thought includes a philosophical bias toward probability theory, direct statistical inferencing, capped with what has been gaining in popularity over the last few years as "rational agents".
But keep in mind - since the 1980's many research cognitive scientists, behaviorist psychologists and neuroscientists have made considerable contributions to the other school of AI - the empirical, sub-symbolic, connectionist view of AI. Take a look at the book Explorations in Parallel Distributed Processing by James L. McClelland and David E. Rumelhart. This latter school is associated with neural networks. Neural networks are more biologically based - generally speaking. In the neural network connectionist based school one finds the philosophical views of Douglas Hofstader, who wrote Gödel, Escher, Bach: An Eternal Golden Braid.
Fuzzy Logic can serve as a "bridge" between these two schools of AI. It has the ability to transform quantitative information into qualitative knowledge and has made a contribution to the computational efforts of managing uncertainty in various business and engineering contexts."
Cog[xii] Robotics Links:
Here's a list of private companies working on robotic projects
Olsted, D. (1998) History And Principles of Neural Network.
Smith, Dr L. (1998) An Introduction to Neural Network
Christian Brothers University directory of research and application sites in neural networking.
KDNuggets: A data mining newsletter:
KDCentral Machine Learning Links
Carl's Research Page referencing his work involving text categorization and multimedia documents
DeAnne DeWitt's career has spanned the gamut from circuit board design to dot.com launcher. Her current incarnation is as a freelance writer, 3d artist and GUI designer. Her portfolios may be seen at www.SpiderFarmer.com and www.ImpetuoCity.com