An Artificial Intelligence Thesis

Warning: this blog is about 12 pages and almost 5000 words long

I finally finished my portion of the manuscript for a book, (tentatively titled Mastering Internet Video), to be published by Addison-Wesley this summer. Thus, I can write about something else without an overwhelming sense of neglect of duty. I've been thinking about Artificial Intelligence.

Impetus

I've recently received my Bachelors degree in Computer Science from UCLA and I've been trying to figure out what I want to do my PhD in. I love teaching and I actually have the goal to start a university in about 10 years, so I figured I'd better get cracking.

As an engineer and inventor, I kept looking around for projects that would be interesting AND useful, to do the research on - I'd like to produce a thing people can use at the end of my Doctorate, not just some inscrutable collection of data added to the obscure corpus of scientific knowledge. I had the opportunity to visit Harvard-Yale-MIT this last summer, and I got all excited about the subject of AI in particular, and started wondering about practical uses of AI and what I could produce. NLP (Natural language processing, i.e. computers understanding human language) was one of the first things I started to think about.

Popular AI Studies

Some of the standard toys to play with in AI are conversation machines (to simulate human conversation), and software "agents" that try to go do stuff for you without being prodded. I've also heard and read unending talk of "neural networks" as an end-all model for brain function and thus the grand unified explanation for how to create human thought or at least an incredible simulation.

The first step of AI is to of course define artificial intelligence, which I define as "acting like humans do when they act intelligently". It's a circular definition because you don't want computers acting like humans do when humans act stupid. You sort of want the "best of" human behavior, which we call intelligent.

I also got a bit of data from my college catalog in the description of "cognitive science" which I guess blends psychology (models of the mind/brain) with computer models of the mind brain. They described that there's two ways to go about creating an "artificial mind": one is to create a model which you believe to be the actual way the human mind works; then, test it in software or hardware. Another approach is to create whatever you want, as long as it simulates the behavior of humans. I thought this was a cool way to break up the problem.

Spiritual Machines

I got most of the way through The Age of Spiritual Machines by Ray Kurzweil which basically said to me that, if you assume that the brain is a supercomputer, since computers are always getting faster, someday we'll probably build a computer faster than the brain. Then of course you can fantasize about downloading your thoughts into the computer, etc. He went on to say that he didn't think it robbed us of our fundamental humanity to be reducible to a computer program, because, I imagine you could sentimentally say, "hey, we're a highly customized computer program based on processing a one-of-a-kind nature and nurture dataset." So in the end we're digital snowflakes.

I remember one time I objected to my friend marveling in a Pavlovian way about how predictable his dog's behavior was, explaining it in this sort of gloating paternalistic way and I got quite annoyed. I realized shortly after that whereas some people object to people being compared to animals, making careful distinction that man is above animals, I was sort of asserting the reverse - that it was demeaning to the dog to compare it to a machine.

The odd thing is, I am probably the most anthropomorphic machine-lover you will ever meet. For years I have called computers "him" or sometimes "her", confusing many clients who looked around trying to figure out who I was referring to when I said "he's confused because you gave him too many commands" or, "he's trying to talk to a different printer..." People eventually became used to the fact that I treated the machines as living entities. My traditional byline is that "I have over a decade of experience making different computers talk to each other" (actually 15 years now - that's from an older resume).

Communication is Therapeutic

As I mentioned in an earlier weblog, WiFi =Computer Telepathy, wireless networking provides frictionless and plentiful communication between devices, which brings delight to technologists.

I've always had a very strong inherent urge to get computers to talk to each other, and it's an amusing thing while technology devices, for many people, are incredible communication tools enabling human to human communication, many geeks are just as content or more so when they get their devices all communicating effectively. When all our devices talk to each other nicely, don't interrupt each other, and communicate high volumes of data, we are happy. So more communication yields happier people, whether it's between them or their devices.

Scientifically Methodical

Immediately after I started sniffing around the field of AI, reading the biographies of AI professors at various colleges I might apply to, I ran into information which ran against the basic operating principles of my own life, which was annoying. I saw a lot of antireligious or mocking speech against religion, which made me think that researching AI would somehow put me in company with hate groups or at least bigots that don't respect the beliefs, religious or otherwise, of others.

It's been popular to make fun of creationists for almost a hundred years now, but someone who would find pleasure in mocking the heartfelt beliefs of others, in my mind hasn't really figured out how to play well with others yet. I began to see why the robots of Sci-Fi always betray their masters and take over the world - because they're created and programmed by know-it-all pricks and we all know that that gets on your nerves after a while, even if you're a robot

I'm a very spiritual person, and so it was glaring to me how much AI research is done by people ranging from agnostic to rabidly atheistic. But, my strong interest in AI means that I must share a lot more with my brothers in this field than I initially cared to admit. So what do we share?

Reductionism

My AI professor at UCLA did a quick sketch on the board and explained a concept of reductionism, where one science is reduced to another. Examples are that biology reduces to chemistry, meaning that all biological processes arguably result and can be explained solely in terms of underlying chemical interactions. And chemistry, it can be argued, reduces to physics, since chemistry is just a (poorly understood) interaction of basic physics particles. The funny thing is, I took a chemistry class at the same time, and learned that chemistry is still using Quantum 1.0 (old quantum theory) to explain its atomic orbitals (which replaced Bohr's orbiting electrons). They haven't upgraded to Quantum 2.0 (quantum mechanics) yet.

I had never really looked at the fact that this reductionism was at work in almost all scientific philosophy. It's a sort of attempt to build everything up from basic principles (like geometry) and unify everything. I was delighted to learn the name for this activity. It can lead to tremendous rounding errors, for lack of a better term, when anthropology reduces to sociology reduces to psychology reduces to biology reduces to physics reduces to quantum theory. Thus, to understand what primitive cultures did, we merely need to grok that quirky quark.

Intellectual Freedom and Paying Back Your Investors

The saving grace of science is, it cares only about results. Thus, whatever the philosophical underpinnings of your principle, if it explains the universe in some novel, demonstrable way, repeatably, then it's science. If you postulate that "matter likes to squish together with other matter, but energy likes to get away from the same kind of energy" and it accurately describes the behavior of gravity and current, then you're well on your way to a scientific principle. If you ascribe human-like emotions to matter and energy in the process, (as in, "nature abhors a vacuum") some scientific purists will point out that the cute and fuzzy characterizations are not inherent to the description of the phenomenon. They may feel compelled to reword your phenomena to ensure that it is cold and unemotional.

On the other hand, just as there have been both religious and atheiststic Existentialists, science doesn't inherently object to a metaphysical worldview. Science doesn't care where you get your appetite for knowledge as long as you come home for dinner.

As long as theory holds water or heats it as the case may be, as long as you can pay back your investors, who cares if your insight came to you in a dream, translated from a mystical language. Or was based on a comic book you read as a kid.

Hidden Assumptions

Hidden assumptions, in my opinion, are major roadblocks to the development of scientific knowledge. Years ago I started writing a short story about a society where the original thinkers had their ability to generate new ideas destroyed by school. It wasn't a morality tale or a conspiracy theory story or anything like that. It was merely making the observation that by training the mind to look at a subject in a certain way, it's possible that this blocks the ability to see the subject plainly, without the (workable but possibly incomplete) thinking constructs created by earlier thinkers on the subject.

The opposite is no solution either- complete outsiders to a subject don't have enough familiarity to innovate or improve. Sure, they haven't been "polluted" by the possibly incorrect conclusions of earlier teachers in the subject... but they can't get results even as good as the earlier experts.

The assumptions don't have to be totally wrong either - they can be slightly not right, or merely incomplete. Einstein didn't refute Newton, he just said that Newton didn't have the right equations for really fast or really small matter. Now, the romantic view of this is that a scientist "questioned the old answers" and somehow valiantly triumphed with his better theories. Sure, we can laugh about it now... usually it's really rough on the ego trying to publish refined truth and you get your teeth kicked in.

Clearly, implicit in my views of science are the ability to get demonstrable results with a theory, without having to sell your theory, use politics to get "buy in" on your theory, use propaganda to disseminate your theory, etc. But science is designed to be above these things, and thus the emperor with no clothes is always eventually disrobed.

Familiarity with the actual thing, and others observations of the thing as opposed to other's conclusions about the thing, are of superior value in researching improved scientific truths. The better a theory can 1) explain existing phenomena and 2) predict new, heretofore unseen phenomena, the "truer" it is in an applied scientific sense.

You Find What You're Looking For

A set of principles, encountered commonly in business motivation and self-help theory, is that "you get what you put there" or "your mind answers the questions you ask it" or "you get what you visualize". These principles are widely regarded as workable, or at least provable inversely, i.e. if you imagine doom and failure, you are more likely to achieve it.

I believe this principle, insofar as it is workable, applies to hidden assumptions as well. The question asked at the beginning of a scientific analysis of some phenomena, if poorly constructed, will insert assumptions into the search and color its outcome.

Example: "Why do humans forget their early childhood?"

This question sets out to analyze a phenomenon, possibly subjective, that humans cannot remember their early childhood. The only logical answer to such a question would be a list of reasons for such forgettingness. If, however, the observation is flawed (perhaps 1% of people do remember their early childhood), then once the analysis is complete, and the standard boilerplate summary of "we're a step closer to understanding why " is published, the statistically true but absolutely false statement, "(all) humans forget their early childhood", gets validated and perhaps other people go on to research the same topic.

Another example of a poorly formulated question would be "why can't metals be transparent?"

Another example might someday be the answer to the question "Why is the speed of light a constant? when some Einstein points out that, yes, it usually is, but when the particles are very ... then the speed of light becomes ..."

Making Problems for Myself

So back to the search for my thesis. I continued to look around, trying to see how I can make some inroads into the very daunting and punishing field of Artificial Intelligence. I remembered that when I first looked at AI as a kid in the late 80's and early 90's, there was a lot of popular excitement about fuzzy logic, about pure AI, about thinking machines that talked to you, fueled by cinema but utterly unachieved by AI science. Funding, both intellectual and financial, dried up and the research had to focus on smaller and more achievable goals, such as language comprehension, stock market analysis, spy photos, optical character recognition, speech recognition. And even these problems tended to surrender, not to a better theoretical understanding of AI, but just by persistently beating against the problems with faster and faster computers. A lot of inventors can look like real heroes when all they do is wait for "brute force", the faster computer, to solve the problem in clumsy but inelegant way. Chess computers may not really play chess like humans; but eventually they can win by just getting fast enough.

As an aside, I recently registered some software with Microsoft, and a robot woman talked me through reading her a dozen groups of numbers and reading me back a dozen more. She, like a human, never repeated the same phrases between groups of numbers. She spoke something like, "read me the first set of numbers", "good, now the second set" "that's great, go ahead and read me the third", "now I�m going to have you read me the fourth..." etc. The point was, this was a fairly simple approach, making a list of many different transition-acknowledgements and randomly (the key to evolving life, we recall ;) selecting one of the phrases to spice up her prose. This itself, was an implementation of some artificial intelligence: to emulate what intelligent humans do.

But following a scripted conversation in a convincing way is not enough AI. Ideally, we'd want to achieve a self aware entity, but we'd happily settle for a creepy and unnerving simulation. The more it fought with us and personified the worst in human frailty, in a convincing way, the more we would feel that we had created life, because only life could act that illogically.

I of course became enamored with one of these unsolvable problems, human cognition. But I've learned in discussion with other PhD graduates that, when their research doesn't pan out, when it essentially fails, they can tell themselves "Well, sometimes knowing what not to research is as good as knowing what to research. This "canary in a coal mine" consolation can be a great euphemism for failure. Because really what you've done is "succeed" by putting a "blind alley" sign prominently in front of your 5 years of research travail for the benefit of others who might accidentally wander that way.

Elevator Pitch

I've worked with enough MBA's to be keenly aware of the need to "exploit a market opportunity". In these terms, I figured I might have some luck approaching cognition in a different direction, one that I would have had to go with anyway because of my fundamental outlook on life.

"How are you different than hundreds of other AI researchers out there?"

So the angle I've thought up is to take a spiritual outlook on life and man and use such an intellectual framework to build up a model of human cognition.

As mentioned above, sometimes the philosophy is superfluous to the explanation. But in other cases, the philosophical framework is a key part of the investigation, a fundamental shift of the basic assumptions, and thus essential to the theory. Case in point:

1. Man is the most cognitively developed animal because of his large brain. He and all other organic life are the result random event where chemicals combined to form life and then evolved through a long span. The unfit or unmutated versions of life died away, but life evolved because of other random processes, because gamma rays or other lucky events not of their choosing happened to mutate them so they survived if the mutation was favorable to survival.
   Man's brain is like a supercomputer. It works differently than our computers in a way we don't totally understand but definitely will eventually. Even if it isn't the same as a super computer, it is similar enough that we can assume it works like a computer, in the sense that it is a data processing machine where you give it a certain (very complicated) input and you get a (very complicated) output. If that doesn't explain all the observed phenomena, remember that random, unplanned events occur in complicated ways (like evolution) and so it makes sense that parts of behavior, or human action we don't totally understand, result from very complicated processes. If you still don't see how super-complex behavior can come from such a simple explanation, think about it in terms of chaos theory: A butterfly merely flapping its wings in one corner of your brain can totally alter your mental weather patterns, so to speak. Because there's nothing inside man's head but the brain, and there is no motion-at-a-distance (ok, except maybe for except for gravity and quantum physics), the thoughts must originate from the brain, probably due to stochastic (statistical) chaotic processes.
2. Man is a spiritual being. The spirit gets injected in the body or breathed into the body or wins the body in a lottery or stumbles into the body or is assigned a body by a bigger, more all-powerful spirit. He animates the body through the brain, which is serves as a routing system for spirit-to-body messages. The spirit can create an effect on the body, and the body can create an effect on the spirit. The relationship is two-way. The animal aspects of the human being are all valid observations; he does have urges and chemicals and glands and basic needs and instincts and passions. But he is also motivated by a large number of goals, conquests, and interactions that are unique to the human spirit and not based solely on stimulus response. On a broad basis, he seeks the conquest of the world around him, and seeks to win a number of contests with and along with his fellow man.
   In terms of mechanics, the body picks up sensory messages and communicates them to the spirit, who could be considered the "black box" that the engineer seeks to study the behavior of. The spirit then puts solutions into play to address the situations presented by the physical world. Understanding the motivations and goals of this spirit, or at least creating a model of these behaviors which generally predicts their activity, could be achieved using any of the hundreds of available metaphysical, spiritual, or religious frameworks available.

A working AI researcher might immediately snicker at this as foolish, or perhaps worse, say that it was no different than what they were already doing. The case could be made that "black boxing" the spirit or "black boxing" the brain achieves the exact same results. But the difference, I argue, comes about from the differing fundamental views (and thus, underlying hidden assumptions and goals) of the scientific inquiry.

Goal Based Machines

I want to ensure that I am completely forthcoming about my complete lack of qualifications in field of AI, having read none of the half a dozen AI books on my shelf, having dropped my AI class after only 3 weeks of "how do I solve dime-store puzzles", and having gotten a C in my algorithms class in college. So, I of course have many years ahead of me of learning what everyone else has done in the field, learning all the blind alleys and conceptual structures of the great AI researchers before me, so that I can hopefully remember why I took it up in the first place and crank out some research and a dissertation in the last couple of years.

That said, I'll briefly list a few of my observations that make me think I might be able to get some tangible results in this field:

• Disambiguation: Pattern recognition, optical character recognition, speech recognition, all depend on a process called disambiguation, where the computer has to figure out why things aren't the same. In general, the computer's job is to find things that are identical, like the children's game of matching the shapes, colors, or patterns. However, logical thought depends on acute differentiation and the observation of how things are similar but not the same. Instead of merely going after identification of items, I will try to program AI that focuses on this observation in its disambiguation system.
• Contradictory goals: A standard theme in morality tales about the evils of artificial life is that the machines have contradictory goals, and that in resolving these contradictions (which humans work at constantly), the computer, with the emotional development of a child, makes some grave mistake and says "I'm sorry Dave" etc., turns on it's creator, blows up the planet, etc. While this is good drama, it seems that we solve this problem all the time in the real world, by raising children with good morals and so forth. Instead of robocopping it and using a "don't violate the prime directive" and giving the AI program a neurosis, I think in my AI research I would work on helping the AI solve it's problems by looking for the source of what was causing them, instead of getting stuck between the two opposed goals.
• Hostile work environment: I've observed ordinarily efficient and productive human staff cower, flail, and make stupid mistakes, because they worked for psychotic jerks. If the AI is supposed to learn from it's mistakes, yet seems to be making a lot of them, I will try to see if I can uncover a source of contradictory goals for the AI. For instance, if a human trying to interact with the AI is naturally a jerk, they will also try to "break" the AI. Thus the AI needs to know that "some people are jerks" and have a system of dealing with this.
• Multidimensional goal set : As we humans haven't yet agreed whether we're motivated by sex, greed, pride, competitiveness, or chemicals, it's understandable why AI is continually given very one-dimensional goals. When AI is given a multidimensional goal, it's usually of the conflicting kind (see above), as some sort of perverse test to see if the AI can resolve the problem any better than it's programmer could. By implementing more spiritually based "meaning of life" programming (any of which tend to be very complex and at least fairly self consistent), a more diverse model of human behavior, with perhaps more depth than "humans were once pack animals thus they like to have friends and also that's why they turn around twice before they lay down on grass..."
• Otherness Racism, brutality and inhumanity stem from a feeling of otherness, that someone else or some other group is somehow not of your own. Forcing an AI construct to "obey it's master" (which backfires) or giving it forced programming to do what it likes but "never do X to a human" (which backfires), creates a behavioral sense of otherness. The sociological implications of mixed race adoption gone awry need to be avoided. We want AI that combines the better elements of humanity and computational power, without wasting processor cycles worrying about indignities it received from humans and how these give it "complex" conflicting goals.
• Emulating human failure: Perhaps the most important aspect of my AI approach is to not go out of my way to make AI emulate the worst parts of humanity in order to make it compatible or more like malfunctioning humans. Mental stress is an emergent behavior of trying circumstances, but the best humans on earth have been renowned for their incredible ability to deal with the most complex human conflicts imaginable. Better to build a "Gandhibot", a "Buddabot", a "MartinLutherKingbot" than a "neurosisbot". I'm not talking about AI based on corny platitudes; I'm talking about emulating the behavior of problem solving, with extremely rich mental models, of universally admired human decision and behavior.
• Avid reading: I think AI should read a lot. Humans continually analyze situations and make decisions, and the applicability and efficacy of those decisions in the real world is determined by the amount of experience that lies behind them. AI does not have the opportunity to be raised by caring parents for 20 or so years, or even to develop for 20 or so years. Thus, it is important that AI at least be fed a large corpus of problems and other people's solutions to those problems. That's not to say that the AI will choose the same solutions, blindly copying what happened in the story. Rather, the AI would translate the stories into concepts, using a straightforward literary plot analysis model like you learn in school (protagonist/antagonist/climax etc.), and use the details as fodder for it's own solution engine.

Artificial Wisdom

As mentioned at the beginning of this treatise, my goal for AI is to emulate humans when they're being their most intelligent, when they're being the wisest. After thinking through what I'd like to accomplish, amusingly I thought I could create a HistorianBot, or an EthicsBot, AI that solved deep practical world issues. It would be quite amusing if you could actually create a AI that could analyze world conflicts, and propose plausible courses of action based on analysis of the various elements. Of course, that would be a sociological model, but the more interesting thing would be if you could create such a system, not by painstakingly modifying the system, but by giving the AI some sound basic principles, and letting it run with them (remember Forrest Gump)

Another way to look at the challenge is this: Imagine if you had to program your friend, perhaps to teach him a new skill or to debug his poor choice in lovers. The challenge, for a friend, would be to instill data and procedures while respecting the dignity and power of choice of your friend. You would try to program them in a way that respected their autonomy. Perhaps AI should be programmed in the same way.

The Age of Cybernetic Spirits

An interesting but more personal question is, how does a spiritual person, believing he is some entity other than the physical world, reconcile an attempt by science to create a copy of him in silicon? Isn't that an inherent conflict with the concept, that life can only come from the combination of a body and some ethereal being that fits on the head of a pin?

The obvious Sci-Fi answer is a "ghost in the machine", the trapping of a soul in some silicon body. Thinking of it in cellular terms, would the spirit split into two spiritlets? Would spiritlets have the same processing power as a single spirit?

A more abstract but more philosophically compatible response is that, the spirit is the source of life. It creates life. Just as it can breathe life into a man, it can breathe life into a place, and activity, and there is no conflict that a spirit could breathe life into a machine. Would the machine be alive? Certainly.

Being a sophisticated carbon-oxygen life form myself, either controlled by a self-reflective carbon-based wet neural-network, or haunted by a self-aware spiritual entity, I am at least partially qualified to analyze human cognition. Whether my intelligence is natural, artificial, accidental, or god-given, I can observe, take notes, and attempt to duplicate phenomena, and failing that I can make jokes about the same.

-dstolarz@robotarmy.com

That's one idea I had. Any other PhD ideas? How will we make robots self aware?