Meg Houston Maker

AI@50: Selected Submitted Papers: Future Possibilities for AI

[AI@50]
Filene Auditorium, Moore Hall
Dartmouth College
Hanover, NH

Conference notes by Meg Houston Maker

Eric Steinhart
Survival as a Digital Ghost

A "digital ghost" is an instantiation of you. [MHM: A machine doppelganger.] The name comes from William Gibson. It's a personalized artificial intelligence, and it could maybe pass a personalized Turing test, convincing your friends and loved ones the machine was you. So it would have to know your history, beliefs, etc.

People now have massive amounts of personal information; there is a digital vapor trail that you leave behind you as you go through the world: videos, images, bookmarks, etc. You go through life generating data, and much of it is not saved or archived. But if you were to start archiving it, you'd get a big temporal database. A little work is being done on this, there's a SIG in the ACM for it, and MicroSoft has its "my life's bits" project.

Think of this collection as a digital diary. Of course you would have to tag all that information -- that's my wife, that's my dog. Tagging and indexing is not there yet -- they're too cumbersome. So instead, you might want a personalized AI that could look at the diary and interpret the diary data as if it were you. The AI tool would have to be a model of your psychology that would interpret it the way YOU would interpret it. Some recomendation engines (Amazon.com) are pretty promising and pretty good at capturing preferences. Mate selection in online data sites could certainly use this technology. E.g. the website Hot or Not, where you rate pictures of people; these could be sent to dating sites.

So, how do you get this? One recommendation is a questionnaire with 20,000 questions you could answer to construct a profile. No way! Too onerous. But maybe we could analyze your telephone conversations, GPS data about where you are at any given time. Or you could do simple query-response about where you were or why you did a certain action. Descriptive and explanatory, in other words. Brain implants are another possibility, or non-invasive brain scanning approaches, to create a highly-personalized analog of your own architecture.

Digital ghosts should also be a simulation of your body. These could be pretty much generic models that could then be tuned with personal history and prefs. Faces and voices are finely tuned personal features, and are interesting to others, so it would be essential to incorporate these. It would have speech synthesizers that could replicate your voice. Your medical data could be incorporated. We can then build a model that looks like you in addition to having your history.

How might we interact wtih something like this? Maybe initially it would be chat-based, but in a more advanced state it would be an animated VR or simulation. It raises a lot of privacy questions, and issues of restricted access.

C. T. A. Schmidt, LeMans and Sorbonne
Did You Lean That "Contraption" Alone with Your Little Sister?

Schmidt's research areas: the dialogical aspects of cognition and communication, and the context for learning -- the physical environment including other people. Key question that interests him: how can the machine learn if it can't communicate?

Robotics-embedded AI should be, or maybe must be, dialogical. In order for advanced humanoid robotics to be fully accepted by others, they will need the proper identity features or they will remain at the fringe of human communities.

Social roles and human institutional involvement seem to have been left out in all forms of AI. To make robots dialogical, we need to work on the pragmatic aspects of communication.

Michael Anderson, U. of Hartford
Susan Leigh Anderson , UCONN
The Status of Machine Ethics: A Report from the AAAI Symposium

Michael is a computer scientist, and Susan is a philosopher. They're here presenting summaries of the papers presented recently at the AAAI Symposium on machine ethics.

The time has come for adding an ethical dimesion to machines -- CareBots, unmanned aircraft, defense uses, etc. This will ensure their actions, especially in self-evolving or learning systems, remain ethical. (Note to contrast this with computer ethics, which concerns hacking and the like.)

The Nature of Machine Ethics:
1) Normative computer agents: computers are normative, because they're designed with a purpose in mind, but not necessarily an ethical purpose. Their performace is assessed according to how well they do what we've told them to.

2) Ethical Impact agents: these not only perform certain tasts, but have an ethical impact on the world. E.g. a robot jockey that guides camels in races, replacing young boys who are slaves who are otherwise forced to do this.

3) Implicit ethical agents: these are machines that are programmed to behave ethically and are designed to perform ethically. E.g. ATMs that are programmed not to cheat the bank or its customers, and automatic pilots entrusted with human safety.

4) Explicit Ethical Agents: machines that are able to calcuate best actions in ethical dilemmas.

5) Autonomous ethical agents: these can calculate the best action in an ethcial dilemma and function independently. E.g., a robo-soldier, sent into battle, which makes ethical decisions that guide its own behavior.

6) Full ethical agents: this term is used to describe human ethical decision makers. Are intentionality, consciousness and free will essential to genuine ethical decision making? Would it be sufficient that machines have "as if it does" versions of these qualities? Could it pass a "Moral Turing Test" for understanding ethics?

If humans create laws that allow them to mistreat entities that resemble human beings, it increases the chances that they will find it easier to mistreat human beings.

Developing an explicit ethical agent is a compelling goal of AI. Many approaches are being pursued. Democracy-dependent algorithms have been created, wherein agents could look up ethical information on the web, giving the machine a kind of "average" or "averaged" ethics. This is probably not good enough. Other methods use neural nets, or offer the human user a case-based reasoning engine with natural language inputs and outputs. Other researchers recommend using deontic and default logics to iteratively construct a theory of ethics. The Andersons have developed a system that extrapolates from experts' intuitions about particular ethical dilemmas.

See machineethics.com

Marcello Guarini, University of Windsor
Computation, Coherence, and Ethical Reasoning

Thagard-Verbeurgt Coherence Theory of Constraints (1998): hypotheses, evidence statements, negative and positive constraints. Used in moral reasoning problems. This system can be encoded in an associative neural network.

Thadarg identifies four types of coherence reasonings contributing to ethical reasoning: explanatory, deducive, deliberative, and analogical coherence. Ethical reasoning is a "multi-coherence" problem. The idea is that it can provide prescriptive or normative recommendations.

People obviously don't arrive at ethical decisions through brute force computation. We don't work out coherence values using internal computation. And we certainly don't do so consciously.

The roll-up: Guarini is critical of the Thagard-Verbeurgt approach, and that coherence is required for moral reasoning in machines. Read his paper for more.

AI@50: Final Poll

40% believe that to produce a successful artifact, the most important approach is to make it interact with the world in which in it operates. Meanwhile, 44% think it should primarily be made to encode and exploit knowledge in the domain

60% strongly agree that AI should take a multidiscipmalary approach, incorporating stats, machine learning, linguistics, computer science, cognitive psychology, philosophy, and biology.

AI@50: Selected Submitted Papers: Future Strategies for AI

[AI@50]
Filene Auditorium, Moore Hall
Dartmouth College
Hanover, NH

Conference notes by Meg Houston Maker

J. Storrs Hall, autogeny.org
Self-improving AI: An Analysis

Self-improvement and self-improving machines: from Turing's "child machines" that would learn and grow, to so-called "stick-built" AI that produces intelligence but doesn't necessarily learn and grow. But human adults do learn and grow, so we need an AI that does that, too.

Universal Learning Machines
There are limits to learning: many animals learn, and have speciific learning programs, but none improve recursively without limit. On the other hand, it's not clear that there are limits to learning for some individuals, and as a species we have been successful, so it's possible that limitless learning is a factor in our collective success. Similarly, our biology is so complex that it's likely we have more capacity. And it also looks like evolution has "learned how to learn;" we didn't evolve to do physics and chemistry, so continued learning was required for these advances.

[MHM editorial: this all seems a bit unrigorous. Is he talking about the simple acquisition of knowledge, or about learning as a process? He's using a lot of analogies to make the case that a universal learning machine is possible. Also, he's conflating the capabilties of the scientific community as a whole, in its capacity to develop a machine that can learn to play chess better than Kasparaov, with the capabilities of an individual human learning agent.]

Selmer Bringsjord
The Logicist Manifesto+

Why the +? You shouldn't just use logic to do AI, you should use logic to discuss ai, too. Logic occupies a unique, indestructibly central position in human cognition.

Types of AI:

- Strong AI: We can engineer systems that have attributes constitutive of human personhood; this is replication, not simulation.
- Weak AI: engineer systems that appear to be intelligent; this is simulation
- Strong/Weak AI: we can manage to create a system that is behaviorially indistinguishable from mere animals and humans, over finite intervals of time.

Strong AI is not likely in 50 years. Weak AI we already have, and Strong/Weak AI, maybe.

How can we ensure that robots always will behave in an ethically correct manner? How can we know ahead of time, via rationales expressed in clear English? We should regulate the behavior of robots wtih computational logic, so that all robotic actions will be sure to be ethical.

How Can We Ensure This?
- Humans select an ethical theory, principles, and rules
- The selection is formalized in logic
- It is mechanized
- Every action that's performed has to be permissible according to this mechanization.

Vincent Muller, American College of Thessaloniki
Is There a Future for AI Without Representation?

Can we do AI without representing the world?

Rodney Brooks suggests a layered subsumption architecture, and situating (embodying or grounding) the machine in the world. He rejects central control over emergence of intelligence. And he rejects representation in the classical sense.

So, what is representation? When should we say that X represents something? What are the conditions for something being a representation? And under what conditions does something represent something else? Representation is dependent on people's intention that something represents something else.

Icons - represent what they represent (e.g. portriat painting)
Indices - connected by a causal connection (e.g. smoke indicates fire)
Symbols - connected to what they represent through convention (like words of a natural language)

Information and Representation: a rock is wet if rainy, warm if sunny. The rock conveys information, but it doesn't represent anything. On the other hand, some things are made to convey informaton for us (clocks, thermometers).

Hilary Putnam, Jerry Fodor and others think that representational or intentional features of symbols and signs must be explained with the help of mental representations. But how do you find out the function (either natural or conventional) of the representation? Causal connection is necessary part of the equation.

Brooks's New AI appears strangely unmotivated -- discovering that classical AI is not represntation, we propose another system that is not representational.

Cognition without central control
Traditional AI/Cog Sci/Philosopy of mind holds that central control is assumed, and then representation is deemed necessary for the agent to succeed.

Theses:
- cognition is the processing of outside information
- the information comes in the form of representations
- the processing of representations is computational

The first two of those make the representational theory of the mind, and all three together yield the computational theory of mind (and the PSS hypothesis).

Indications against central control:
Mark Bickhard: critique of "encodingism:" - the mind does not decode and encode information: because who would do the decoding and encoding? Ways out of this: remove the notion of encoding from that of representation via natural functions, and abandon the notion of mental representation as a causal factor.

John Searle's Chinese Room Argument
Searle looked at the processor in a computer and asked what that agent understood: nothing! The systems reply agrees that the agent does not understand, but claims that the whole system, of which it is a part, might. Buty why should the system understand?

Brooks's inspried response: the system understands because there is no central agent in the system.
Abandon the cartesian theatre: searle was asking the wrotn question, just like eerybody else.

Embodiment of cognition: intelligences in the body as a whole. This involves the integration of cognition and action. Perceiving isi really a way of acting. Perception is not something that happens to us (see e.g. Noe). So there is an indication that cognition, emotion, and volition must be integrated.

We do not get conscious representation without central control. Conscious experience is the experience of "what it is like" -- the having of experience: I am the one that does the perceiving, thinking, palnning, and acting. Consciousness requries something being conscious, and this requires a central agent. The progelm is people move from that to saying that all of cognition has to be centrally organcized. But we don't know what we can do (or do without) central control.

AI@50: Future Interactions with Intelligent Machines

[AI@50]
Filene Auditorium, Moore Hall
Dartmouth College
Hanover, NH

Conference notes by Meg Houston Maker

Daniela Rus, MIT
Making Bodies Smart

Robotics is a dream to build a machine in your ownd image; building platforms for stuying life and intelligence, intelligent conection of perception and action, outcome of visiouaries with energy and resources; the next disruptive technology. The field has been fueled by crazy dreamers.

The history of robotics is the history of automata, Golems, Frankensteins, all the way to current robotics. There has been some trepidation about these "soulless beings."

The hard problem is to capture what happens in the interaction between computation and the physical world. Hand-eye coordination, which we do to survive, is really one of the hardest problems.

Robotics methodology timeline:
- Knowledge representation -> symbolic reasoning
- Blocks world - kinematics
- Motion planning - automation and manufacturing
- Reactive control - exploration
- Probabilistic/expectation/verification - mobile/field, space, underwater
- Human-centered - home, medical, entertainment
- Modular/distributed systems - self-organization, RoboCup

Things we take for granted in robotics (and need to do our work): GPS, Real-time vision and stereo, WiFi, Rapid prototyping, GHz processors, Gumstix, Sick laster rangefindesrs, etc. These have all developed in the last 10 years, and have been a boon.

What we've gained: localization, planning and navigation, 3D perception. However, we have many challenges ahead. Of these, uncertainty is one of the biggest; the world is unpredictable, and the robot has to be able to handle it. Perception also remains a major challenge. Robotics has made huge strides in areas of navigation, but manipulation, which is just as important as navigation, is far behind. We need to be able to develop algorithms for enabling computers to interact better with humans, and we need deeper connections between perception and action.

"Simulations are always doomed to succeed:" taking the robots from the lab to the field, from simulation to actual world, remains a big challenge.

Many robots have not been constructed with the explicit goal of understanding human intelligence. Rather, they have been more task- and motion- oriented. What is the environment of the robot? Is it fully known? What is available to it? This is robotics for structured environments, and comprises the last few items on the list above (motion).

Rus shows her work in with robotic self-organization; really smart parts that know how the whole is to be organized. In application, imagine a robot encountering a river without a bridge, and constructing a bridge for itself.

Predictions for the future:
- Robotics will give us new theoris for intelligence -- we need to account for situatedness and embedding in natural world; we will move beyond Turing intelligence
- Robots/actuation will be pervasive
- Robots will accomplish the ultimate automation -- on-demand factories
- Robots will save the environment -- biologists are overwhelmed by simple tasks like data collection; it's not just to monitor environment, but also to effectively repair processes in which the system is damaged
- Robots will extend and enhance life -- the convergence of DNA and silicon will happen.

Human-centered robotics will enable us to move from personal computers to personal robots.

Sherry Turkle, MIT
From Building Intelligences to Nurturing Sensibilities

Technologies are never just tools; technologies are "evocative objects." Computational entities (computers, agents, robots) become relational companions. They provoke anthromorphizing. Her objects are designed to provoke nurturing behavior.

What is a relational artifact? It's an evocative object. These present themselves as having states of mind affected by interactions with humans. They impress not just through their intelligence, but through their sensibilities. In how we interact with them and reflect on our relationship to them, we learn about our own inner states.

E.g. Kismet, Aibo, Furby, My Real Baby. This last wasn't so successful with children but was very successful with seniors in senior centers. Children and seniors are both somewhat vulnerable populations. With seniors, sometimes it's interactions with robotics or nothing, and in some cases the robot is "good enough" to satisfice, as it were, for human interaction. But the robot is not really empathetic. This could be misused, morally, because the question is who will be given (actively) a robot as the best our culture can do to provide them with companionship: the disabled, the elderly, children.

The art and sciene of reltional artifacts is part of the art and science of understanding human psychology and human vulnerability. Interaction wtih these objects move from the psychology of project to the psychology of engagement and object relations. In talking about relationsl artifacts, the discourse of aliveness moves from the objects competetency (physical or psychological) to the person's relationship with the object. They push our Darwinian buttons. People respond to these objects as though they were sentient and feeling objects.

Children can differentiate -- and have to learn to differentiate -- between animal alive-ness and robotic alive-ness, between animal love and robotic love. Children nurture robots; they think of a nonfunctioning robot as wounded, they show belligerence towards robots when angry, and view robots as dialogue partners in conversation.

Nurturance may turn out to be the killer app.

In the 1980s Turkle met a teenager who spoke about the pleasure of working with a personal computer because it was like putting a piece of her mind into the computer's mind. Twenty years later, children and elder nursing home residents who have robotic agents report that they feel emotionally connected with the robot. They feel love. This is a sea change.

Relational artifacts present themselves as a possible Self objects. Now, Self objects are positive psychology when the object is a person. It presents a narcissistic experience to project self onto others, but robotics might be used this way. Perhaps there is therapeutic potential here, but what are we to make of this? Is it good for us that we have robots that convince us they care? Or will our lives be enhanced because we'll have to develop new ways of feeling toward these others, and our new kinds of relationships?

AI@50: Sixth Polling Question

Only about 10% agree with Marvin Minsky that AI chess is waste of time because it's about exploiting human weakness, not simulating human thinking

AI@50: AI and Games

[AI@50]
Filene Auditorium, Moore Hall
Dartmouth College
Hanover, NH

Conference notes by Meg Houston Maker

James Moor, intro: Chess was thought to be difficult, so that if one could master a chess program one had achieved something. However, it turns out games are reasonably easy to master, and areas like natural language are much harder. Moor plays a video clip of Marvin Minsky, who says that chess AI projects had to be focused on the middle game, and the opening had to be programmed, so AI chess was not a very promising simulation of human intelligence.

Jonathan Schaeffer, Alberta
Games as a Test-bed for Artificial Intelligence Research

Marvin Minsky is WRONG WRONG WRONG! But he's not even here today to hear me say it! [Minsky left the conference yesterday.]

Building a chess program was viewed as a grand challenge of AI research. It was a goal that seemed attainable, and worthwhile, and was something that the public could relate to (unlike missle trajectories). This was a coup for AI, because this made it accessible to the media. The first computer chess tournaments started in 1970. So this is really the longest-running experiement in AI, and we can see the evolution of successes. Also, chess products were AI's first big commercial success.

Schaeffer clarifies he's not a games researcher -- he's an AI researcher who uses games as his testbed, rather than defense or engineering proglems, etc. He takes an application and does whatever it takes to build a powerful and successful program. It's an overall strategy of building a high-performance system.

In the chess world, we have seen lots of effort in the last 40-50 years. There was a lot of effort that went into building a grand master chess program. It took a lot of effort and money to build it. Visionaries include Claude Shannon, Alan Turing, John McCarthy, Newell and Simon. They underestimated the challenge!

But it turns out that to make a winning chess program, you need an Alpha-beta search algorithm, an evaluation function to decide good and bad chess positions (these are pretty simple), and fast computing hardware. Wtih hindshigt, this all seems obvious, but it was not evident until the 1980s. Hardware was the key to Deep Blue's success. It had 512 special "chess chips" running in parallel. It was a very fast parallel process. But it's an idiot-savant: it plays brilliant chess, but it can't do anything else.

Brute-force search: use a comptuer to do fast searching. Many researchers felt this was not actually AII because it's not how a human would play chess. But there's no argument that can be made that says humans play chess the "right" way. Computer hardware is fundamentally different from hnuman hardwared, so it shouldn't be surprising that these machines operate differently in solving the same problem.

So one of the most profound contributions of chess is the effectiveness of brute force search. It's simple, elegant, and a good match fo the capabilities of the computer. Other AI chess contributions include search enhancements, hardwares colutions, and an appreciation for the difficulty of building high-performance AI systems.

Chess itself was not as interesting for AI researcher (limited, 2-player game) as other games, which were more representative of the real world, involving randomness, etc.

Checkers: Arthur Samuel used checkers for pioneering AI research. This work, from the 1950s and 1960s, is still cited today. Marion Tinsley was a checkers grand master who only lost 3 games in 40 years, but he was beaten in 1994 by Schaeffer's team developed. This system earned the right to play Tinsley by winning up through competition.

Other games: Othello is another game in which the AI machine version completely overwhelmed the human chamption. Backgammon and TD-gammon, a neural-net program, showed other successes. Scrabble also fell (1998), using Brian Sheppard's Maven. To a computer, words are just sequences of letters. Crossword puzzles (1999): Michael Littman and a team built Proverb by building a database of past crossword puzzles. It turns out about 1/3 of clues are repeats from previous puzzles. So just looking up clues can get you 30% performance. If you take the clues and send them to an internet search engine, you get the result pretty quickly. The program does 96% on the NYTimes Crossword Puzzle, without understanding the meanings (semantics) of the clues at all. Poker, Bridge, and Shogi AI products are coming soon. Go still needs a lot of work.

Games are great becuase you have no trouble attracting grad students!

But chess is the past. So what's going on now? Commercial computer games are now a $25 billion a year industry. This isthe new grand challenge in using games as an experimental testbed. Interactive games are the killer app for human-level AI (see John Laird and Michael van Lent).

Example game: Oblivion. It boasts "radiant AI" with realistic characters. Schaeffer plays a clip that is laughably unrealistic. Characters don't react in naturalistic ways to actions in the game.

Why is Game AI Hard?
Games have to be developed to run on a 2-year-old computer, and can't use the whole CPU (just 10-20%). You have to use limited memory and limited disk, and the program has to be thoroughly tested. You need real-time response (no long computing times). Content generation is the dominant cost in R&D. Realism is essential, and the game has to be fun. The Academic AI community doesn't develop solutions useful in the game AI community. Researchers think nothing of using lots of memory, long processing times, etc.

Random numbers are the most powerful AI tool. Because from a random sequence of events, users discern intelligent behavior. It's not truly random, it's controlled; it's not as if the character walks randomly choosing directions. Rather, it's controlled in that having made a random decision, the character's next decision is consistent with that first decision. So it looks like flow, it looks naturalistic to users.

Obstacles to Quality AI in Gaming
There is little research money in industry, and the game industry is suspicious of academic research: about 90% of what's published as "useful for games" is not actually usable by the gaming AI community.

Want a challenge?
Get rid of your fast hardware, use little time and memory, try to create realism and give real-time response. Human-level AI is needed today.

Danny Kopec, Brooklyn
Chess and AI

What you see in matches between humans and computers relies on human frailties. Machines can learn from humans, and humans can learn from machines.

Chess is very much alive. It hasn't really been solved yet. In May 1997, IBM's Deeper Blue defeated Kasparov, but Kopec doesn't put much stock into this, because it was a very quick match, and Kasparov changed his usual strategy, and the machine exploited that. Kopec gives many other examples of chess matches between humans and computers over the last 5-10 years. Results are mixed, but adding complexity is that the tournaments and exhibitions are being run often with times controls -- 60- or 90-minute time control (match length) with a 10- or 30-second increment (time to make a move).

Is the machine "thinking" when it's calculating? There is disagreement here. There's nothing to say that they have to "think" the same as humans. Kopec shows many examples in which the right play is counterintuitive. Humans play by heuristics, experience, knowledge, intuition, knowledge, and pattern matching. Human chess involves psychology: gambling your own knowledge against your opponent's knowledge. Humans search what's important on the board, and know how to choose what's important. Chess is way more than search.

AI@50: The Future of the Future

[AI@50]
Filene Auditorium, Moore Hall
Dartmouth College
Hanover, NH

Conference notes by Meg Houston Maker

Ray Kurzweil, Kurzweil Technologies, Inc.
Why We Can Be Confident of Turing Test Capability Within a Quarter Century

Kurzweil demonstrates some new developments: he takes a picture of a page of his recent book, then in a few seconds we hear the text read through his laptop. This is not your daddy's OCR and text-to-speech system; it is managing the natural image distortions that arise when you take a picture of a book with a hand-held camera.

The pace of progress is not constant. The paradigm shift rate is now doubling every decade. The pace of technological evolution and of technological adoption is accelerating. Exponential systems go on until that particular paradigm runs out of steam. And then we're ready for the next paradigm.

Moore's Law was actually the fifth paradigm for computing; right before that there were vacuum tubes, which were shrinking as capacity grew exponentially. Then we got chips, and pretty soon we won't be able to get further exponential growth from our existing silicone architecture and we'll have to move to 3-dimensional chips. Intel's pretty confident these will be ready so that capacity will continue to follow that exponential trajectory predicted by Moore's Law.

But, what are the limits of growth? In computation, it looks like there are limits, but it turns out theyre not very limiting to what we're trying to do in AI. By about 2020, we'll have enough computational power to simulate all the features of the human brain. So what about the software? Some think that reverse engineering the brain will provide the ultimate source of the template of intelligence. Our brain scanning software is getting more powerful, with more resolution, each year.

What is the complexity of the human brain? There are trillions of interconnections. But there are only a few genes that control it; the design is in the genome, and the genome just doesn't have that much information in it (it has a lot of redundancies). It's managable complexity.

AI@50: The Future of Language and Cognition

[AI@50]
Filene Auditorium, Moore Hall
Dartmouth College
Hanover, NH

Conference notes by Meg Houston Maker

Trenchard More, IBM (retired)
The Birth of Array Theory and NIAL

More dives into a discussion of both simple and elaborate Cartesian products of arrays and array nesting. I bet it's beautiful to a mathematician. I can't possibly summarize it. See more on Array Theory and NIAL (Nested Interactive Array Language).

Eugene Charniak, Brown
Why Natual Language Processing is Now Statistical Natural Language Processing

NLP has experienced a revolution in the last 15 years, and gone from being a domain based on logic to a domain based only on statistics. Non-statistical approaches are now quite rare.

Parsing
Going from a string of words to a parst tree. "Alice ate yellow squash" gets decomposed to noun, noun phrases, verbs, etc., in a complete analysis of the sentence in tree structure. This is much like what a grammar school student does when diagramming sentences. Parsing gives us a first pass at a meaning of a sentence, breaking up the meaning of the whole sentence into the meanings of its parts. But if you don't know the rules of the language, or grammar, you're not going to be able to get a correct parsing or correct interpretation.

Adding complexity, a single string can have multiple parses. So, which is the correct parse? Different parses yield different meanings:

"The salesman sold the dog biscuits."

Did he sell dog biscuits?
Did he sell biscuits to the dog?
Did he sell a dog named Biscuits?

Parsing in the "old days" assumed that sentences had only a handful of correct parses. Most research was concerned with either designing a parsing algorithm or developing better algorithms. As late as 1990 there was no parser that could find a semi-reasonable parse for every sentence on, e.g., the front page of today's New York Times. Which is a perfectly reasonable requirement.

So why is this problem so hard? First, grammars "leak" -- it's hard to write a grammar that handles all the complexities that natural language can throw at you. The second problem is ambiguity. To prevent a grammar from leaking, you have to write a grammar parser that is overly complex (and unrealistic).

We have to recharacterize the problem as a probablisitc/statistical one to find the most probable parse for a sentence. Probablistic context-free grammar offers a set of rules that define derivations for strings in the language of the grammar, with derivations represented by trees. The probabilities are then interpreted.

Training corpus: The Penn Wall Street Journal Tree-bank, which contains about 1 million words, and was compiled by humans. Lexical statistical parsing is achieving decent accuracy, and very promising future results.

But still... how to get at meaning? Take it gradually. Start with the stuff you know. E.g., you know that if a sentence has "tomorrow," that's it signifies temporal meaning.

Pat Langley, Stanford
Intelligent Behavior in Humans and Machines

We need a renewed emphasis on cognitive psychology. Simon and Newell, the early pioneers, viewed themselves as psychologists trying to explain human thought. This approach was well respected at that time. Much other early work on representation dealt with the structure and organization of human knowledge. Not all research in AI was motivated by psychological concerns, but it had a strong impact on the field.

There were studies of human problem solving in early AI work and these led to key insights about both state-space and goal-directed heuristic search. The 1980s saw many developmetns in knowledge-based reasoning that incorporated ideas from psych: expert systems, model-based reasoning, analogical reasoning, and natual language work.

Early AI modeled human learning and discovery, problem solving, categorization, natural language, and scientific discovery. Modern AI has moved away from human cognition and is unfamiliar with results from psychology, despite the historical benefits. E.g., in knowledge representation, the work focuses on efficient processing. In problem solving and planning, and NLP, there are focuses on statiscial methods that have few links to linguistics or psycho-linguistics. And in machine learning, statistical techniques are dominant, but you need larger data sets and training cases than a human would need to achieve the same level of expertise.

Technhology is one reason for this shift. We have faster computers that allow us to mine large data sets for more predictive models. But we should combine this computing power with insights from psychology. Additionally, AI is typically housed in computer science departments, which often grew out of math departments. Mathematicians are usually more concerned with tractability, even if that restricts the work to narrow classes of problems, and many mathematicians also view psychology with some skepticism. There has been commercial success with AI, but this tends to encourage niche works. In academic settings, we should reach for higher goals.

One reason to renew the interchange between the two fields is to achieve better understanding of human cognition. Human abilities serve as important sources of tasks for study: studies show that people can do some pretty impressive things.

Cognitive psychology makes important representational claims, e.g. that humans resort to problem solving and search to solve novel and unfamiliar problems, and that they use heuristics (not algorithms) to find good solutions. In learning, cognitive psych shows us that people create new cognitive structures in problem solving, and that learning is incremental and interleaved with perfomance.

The AI curriculum should include cognitive psychology, structural linguistics, logical reasoning, and the philosophy of mind.

AI@50: Sixth Polling Question

[AI@50]
40% say it will be more than 50 years before machines will be able to simulate learning and other aspects of human intelligence

AI@50: The Future of Reasoning

[AI@50]
Filene Auditorium, Moore Hall
Dartmouth College
Hanover, NH

Conference notes by Meg Houston Maker

Alan Bundy, University of Edinburgh
Constructing, Selecting, and Repairing Representations of Knowledge

Two hypotheses
1) The representation of a problem is often the key to its solution
2) Problem representations can be automatically formed and automatically repaired

Representation as the Key to Problem Solving
- McCarthy's mutilated checkerboard example
- Saul Amarels Missiories and Cannibals example -- illustrating how a change of representataion affects the size of the search space, making the problem easier to solve
- Andy DeSessa's Bouncing Ball, in which it's asked: where does the energy of a bouncing ball go at the moment of impact? The answer is that the ball and floor both become deformed. It is essential to understsand that neither is fixed.

We have to think of representation as fluid, as something that changes as we reason. The world is infinitely rich, and we cannot model every aspect. Fixed representations cannot cope with a changing world and new challenges. If we want to build agents that can deal with a changing world, then we'll have to have representations that are under machine control and will evolve according to new situations.

This is not a question of changing your beliefs about the world. Rather, it's a change in the "signature" -- the predicates and functions in your algorithms. You might need new functions, you might change their arguments, their types, etc. Even the logic might have to change. And all this has to be under machine control during the processing of problems. Computers can form representations and then be made to repair faulty representations.

Repair of Representations
Cynthia, e.g., is a program that helps users write computer programs, creating new concepts and conjectrures from examples. Underlying the system is a proof that the current program was not meeting expectations of well-formedness. Other systems work to identify faulty proofs that end up causing bugs in the system. The ORS Program, e.g., repairs faulty ontologies by analyzing failed multi-agent plans. Changes include abstraction and refinement of signatures by adding arguments, changing predicates, etc.

This technology is essential for realizing the Semantic Web, in which huge numbers of agents are going to have to share ontologies. The changing community of systems will never really share any single ontology -- there will be sublte differences no matter what we do. The idea is that one can dynamically repair those differences as a result of faulty interactions between the systems. It's not possible to do static repair when you have huge numbers of agents independently developed.

Conclusion
- The formation of representation must be under machine control in order to deal with a mutli-agent, changing world.
- Representational change can be triggered by reasoning failures

Edwina Rissland, UMass
The Exquisite Centrality of Examples

The world is rich, it's got exceptions, it's not regular, and yet we can deal with it. Rissland calls this the inevitable intertwining of knowledge -- a Yin and Yang intertwingling of (Examples) with (Concepts, theorems, models, statistics).

Examples are grist for the mill of learning. Rissland's interest is in the Example side of the equation:

- in learning and in teaching
- in explanation and argument
- in concept representation and concept change
- examples built to suit; in how you build examples that satisfy the problem you're trying to solve

Categories of examples (or hypotetheticals, which are equivalent):
- startups, which are easily used without much background
- references, which are standard or useful anytime
- models or prototypes that capture generalites or abstractions about the world
- anomalies that are atypical or borderline
- counter-examples used to refute or limit

There is structure implicit in probing with hypothetical examples.

Examples are useful when we're dealing with real world concepts. Our analysis shoudln't be all representation, or all examples, or all statistics. She advocates a hybrid approach: open-textured (messy) + non-convex + non-stationary

Conclusion:
Examples are really halfway between statistics and generalities. Don't throw the examples away as you try to abstract up. Examples are useful in concert with other information (from the domain), plus logic, statistics, etc.

Q&A
Q: Wouldn't it be useful to have a knowledge base of examples?
A: Some fields do: medical bases, legal case books, etc. These tend to be domain-specific. The trick is how to represent these and put them into your knowledge base. Building examples with constrained example generators works when you know what you want. But how do you know what you want?

Q: Are examples mostly concrete instances?
A: They don't have to be. Once you start developing hypotheticals, you start expanding beyond examples to logical models (at the extreme).

Bart Selman, Cornell
The Quest for Machine Reasoning

From the 1960s to the 1990s, computational limitations made machine reasoning nearly impossible. McCarthy (1956-59) proposed an intelligent system architecture with declarative representation and inference. But inference was computationally infeasible. From car repair (with about 100 variables) to military logistics (with 200,000-600,000 variables) and more, the computational requirements are enormous.

But in the last 15 years we've made progress in processing in these combinatorial search spaces. Applications include pattern generation, e-commerce and e-auctions, scheduling, hardware and software verification, multi-agent systems, etc. Significant advances have brought us from being able to handle 200 variables to 1 million+ variables.

Selman shows a chart showing the time required to solve specific reasoning problems, using technology available over the last decades. Now, the time to solve some of these classic problems has been driven to zero. (Someone interrupts to ask what zero means. The response: "That's very fast" [laughter].)

Backdoor variables: real-world structure hidden in the networks can be exploited by automated reasoning engines. You can reduce a many-hundred-variable information space by isolating a few backdoor variables that will solve the problem.

Summary
Automated reasoning tools might serve as a true "cognitive assistant" helping humans apply their creativity to a system, and use the solvers to analyze the implications of these choices, and for validation and feedback.

AI@50: Fifth Polling Question

[AI@50]
41% strongly agree and 44% somewhat agree that vision shouldn't be treated as input chanels converted to symbolic representations; instead, we should assume there might be intelligence going on earlier in the system

AI@50: The Future of Vision

[AI@50]
Filene Auditorium, Moore Hall
Dartmouth College
Hanover, NH

Conference notes by Meg Houston Maker

Eric Grimson, MIT
Intelligent Medical Image Analysis: Computer Assisted Surgery and Disease Monitoring

Vision is primarily an inverse problem. The goal is to infer properties of objects in the world to support intelligent interactions. You want to recover some or all of the underlying information of the image.

Historical approaches are restricted to low-level processing: edge detection and shape recovery. We tended to ignore surface texture, etc. So you could have a representation of the geometry of the object, but then you have search problem, matching the detected results against a set of knowns. Is there a recognition of a particular geometric shape within what's detected? This doesn't work well in cluttered scenes -- selection, grouping, and focus of attention are critical to controlling complexity.

Recognition of shapes not well described by geometry is elusive. E.g., why can't you just recognize a cat? Well, if you could render a cat as polyhedrals, you could do a good job recognizing the cat. Many approaches were originally driven by the need to control complexity due to computational restrictions.

There has been a shift; we're now using more statistical methods for object recognition. E.g., SIFT methods (Lowe) or constellation methods (Malik, Perona); but these are mostly just more complex edge-detection. But we now have learning methods that allow us to train the system.

Examples from medical image analysis and image guided surgery: AI assists in building accurate models of the patient (e.g., the tumor in the patient's brain). That model is then used while conducting the surgery. The model takes input from MRIs or similary patient scans and makes inferences about the structures it sees (gray matter, white matter, tumor, etc.) Statistical learning is a big part of these methods. Domain -specific knowledge is incorporated in a way that reflect the details of the domain. The system outperforms the radiologist on specific tasks.

Statistical learning methods will continue to be central elements of computer vision approaches. A critical problem is incorporating prior knowledge or expectations -- learning where to expect different types of structures or features in the object.

Q&A; Grimson's comments:
So, we still can't recognize a cat, but we are closer. A good vision system would actually watch the cat walking, and do some math to figure out it is a cat.

The medical work is high-stakes work. So we work with the surgeons closely and build-in error monitoring to alert them to likely modes of failure.

Takeo Kanade, Carnegie Mellon
Artificial Intelligence Vision: Progress and Non-progress

Kanade developed a Matrix-like replay system called EyeVision for the Super Bowl 2001. Each of 30 cameras tracks an event with zoom, focus, and tilt in real-time. The system comprised 30 cameras mounted around the upper decks in the stadium in Tampa FL. Each camera was controlled remotely and the composite allowed for 3-d reconstructions of the actual footage. With one camera, you can't see the hole for a quarterback to make a pass. But when you swing it around in replay, you can see a big hole for him to pass to the receiver.

Vision is difficult. Vision was one of the earliest problems in AI: blocks world, outdoor scene analysis (which was overly ambitious when we look back on it today). Why is it so difficult? Lots of data: in one 1000 x 1000 image, it contains 10-6 10 x 10 images. In a 10 x 10 block at 8-bits per pixel, the number of possible images is 256-10 x 10 which equals 10-240. So, his theorem:

"Human kind has not seen all the 10x10 images yet." [laughter]

Context is another key difficulty. Why is that an image of a car? Because it's on a road. Why is that a road? Because there's a car on it!

We need a more disciplined approach for searching and defining constraints:
- physical and 3D recovery (shape from X methods)
- geometrical reasoning
- statistical modeling and learning -- statistical properties are quite often enough to recognize an object

He's developed a method of, essentially, shaking the head (or vision system) just slightly to detect the ground layer and obstacles. Edge and color detection produces useless and noisy output, while this produces correct structured images of the meaningful objects in a scene, and allows the disambiguation of figure from context or background.

Vision is physics, geometry, and statistics; this appraoch is generating real results.

Q&A
The fact that animals that move their heads to detect objects stereoscopically is interesting. It's great that this is in the real world, but the really interesting thing is that this is a rigorous method that works.

Terry Sejnowski, Salk Institute, UC San Diego
A Critique of Pure Vision

His goal: To understand how the brain works. But how will we know if we have the right principles, and if we really understand it? In his view, if you really understand it, then you can build it.

Sejnowski once asked Allen Newell why Newell had turned his back on neuroscience. Newell replied that at that time the field was too young to provide any useful information. That's all changed now, because we have much better techniques for analyzing the brain at all levels of investigation -- from the molecular level to the level of large structures.

Pure Vision: the idea of a feed-forward network that allows us to construct an image of the world. As we saw in Tadeo's talk, that if you have a motor system, you can do a few more interesting things.

In "A Critique of Pure Vision," Sejnowski and others argued there is weak evidence for the basic assumption that vision is a feed-foward system. Change blindness is an excellent argument against this approach. The test to show this consists of blinking an image on a screen in which one object in the scene is coming and going; observers have a hard time figuring out what part of the scene is changing. There's so much in the scene to pay attention to, that there's no way to keep track of it all at once. We now know what mechanisms in the brain allow this to happen; it's an internal neurological phenomenon, and attention modulates the firing rate fo the neurons in the recognition apparatus. So this is activity in an internal signal -- it's not coming in from the outside world.

Sejnowski cites many animal experiments that show temporal difference learning. The take-away is that techniques for studying neurons and reactions to stimuli are key to our understanding the cerebral cortext and the computations it's doing.

We don't know what the brain is doing. The only thing we have is the end of the computation -- what pops into our head. But all the heavy lifting is happening beneath the surface. We need to go beyond the surfaces and find the algorithms being used. Temporal difference learning is just one of these.

Q&A
Q: There's not a lot of clarity about what's AI, what's cognitive science, what's neurscience. What you're doing is not obviously AI.
A: My goal is to understand how the brain works. If all you have is the input and output, you're doomed. You need a way of understanding how the whole system is organized in between.

AI@50: Fourth Polling Question

[AI@50]
30% strongly agree, 31% somewhat agree, and 27% somewhat disagree that the future of AI will be propositional or probabilistic vs. based on complex knowledge representations
38% somewhat agree with brute force solutions vs. modeling of human thought
36% strongly agree and 43% somewhat agree that successful development will combine machine learning with automation of deductive reasoning

AI@50: The Future of AI

[AI@50]
Filene Auditorium, Moore Hall
Dartmouth College
Hanover, NH

Conference notes by Meg Houston Maker

Rod Brooks, MIT
Intelligence and Bodies

Brooks uses a Mac

Cybernetics: 1948-1961: Control and communication in animals and machines, including the study of all possible behaviors of a machine, and modeling animals as machines. By 1952, Asby advocated considering the animal coupled with the environment. These researchers built instantiations, but didn't use digital computers (because they weren't invented yet). Grey Walter built two robot "tortioses," Elsie and Elmo, and published articles in 1951 showing these robots interacting with the physical world, responding to light and touch.

So the key "Dartmouth Innovations" were:

- Modularity: an element of abstraction
- Ungrounded symbols (another element of abstraction): put the human observer in the role of connecting the program to the world; let researchers jump to much "higher level problems"
- Search: Turing, Michie, Shannon, and Weiner had all considered this for chess
- Jumping onto the digital train as the mechanism for expression of intelligence (vs. analog circuits)

"The core of our humanity, our abillity to think and reason, was subject to our technological understanding."

CS and AI have largely merged over 50 years. The first technical approaches with in the new framework got replaced later by new innovations. Initial "problems " ended up being category errors. There is an inconvenient truth that applications drive science, despite scientists' distaste for this fact.

The AI founders -- Turing, Michie, and others -- were mathematicians and were able to reason about logic problems, theorems, non-chance games. This is what intelligence was to them. They were smart! So they weren't as interested in solving "simpler" intelligence problems [like how you manage to walk downstairs and get yourself a cup of coffee].

LISP gave us recursion and manpulation of ungrounded symbols (at a time when computer architectures used intrinsically non-recursive call structures). It gave you a control stack. So search routines became easier to write than they were in Fortran, and symbols became easier to use because they were pre-parsed. So AI became search- and symbol-centric because of the availability of this tool.

However, all Brooks's robots use LISP! Including the vacuum cleaners -- and there are about 2 million of these out there right now running LISP.

If you review how the brain was thought about over history, it's always been described in metaphorical terms. In 1900 it was "photographic," in the 1950s it was "a telephone switching system" in 2000 it's like the "world wide web."

Remember: descriptive is not generative, and the system is not the model.

Embodied AI: Robots in the World
Brooks shows clips of his robots interacting with objects, reacting to objects handed to them. They produce action in the world, without evident planning. Brooks argues that planning is an emergent property of the system. It's like designing a bridge -- you build the bridge to be very strong, but you don't put a "strength module" in the bridge.

The Future of AI
As Boomers age, this demographic trend is going to push for an increase in physical services -- more automation, more help with the physical world. AI research should focus on developing agents with:

- the visual object recognition of a 2-year-old
- the manual dexterity of a 6-year-old
- the language ability of 4-year-old
- the social sophisitication of 10-year-old

Nils Nilsson, Stanford
Routes to the Summit

The "summit" is human-level artificial intelligence. There are probably some false summits. There may be more than one summit. There is no single route to the summit -- search, reasoning, reinforcement learning -- will all end up being necessary and incorporated in reaching the summit.

Early Pioneers: Turing thought computation could include everything human brains do, no matter how creative or original. The McCulloch-Pitts Model of Nerve Nets were forerunners of the "sub-symbolic" AI approach. And the PSS: a physical symbol system has the necessary and sufficient means for intelligent action (Newell and Simon).

PSS = Formal Symbol Manipuation + designation and interpretation
An expression designates an object if the system can affect the object (action) or behave in ways dependent on the object (perception)
Interpretation is running programs
So PSS doesn't necessarily preclude connection

Promising Approaches:
GOFAI Approach: AI programming looks at a specific domain (geology, medicine, chess, chemistry, etc.). But these approaches are brittle and domain-specific, so you need to incorporate common sense into the system.

Cognitive Substrate Approach
Build a core or cognitive substrate within the system that has the abilities of a child. We need to program enough of them so that through learning, education, training, etc., the system could eventually become expert in a field (geology, medicine, chess, chemistry, etc.)

Computational Models of the Neocortex
Bayes network structures, both shallow and deep nets: simple inputs are fed up the hierarchy to increasingly abstract representations of the sensory inputs. This is promising, becuase we don't just react to raw sensory data; we also (only?) react to abstractions of data.

Pandemonium Reprise
Selfridge's Pandemonium consisted of cognitive demons, computational demons, and data or image demons. The system was both hierarchical and parallel; weights were adjusted by hill-climbing. Demons were committed neither to symbols nor to neural nets; it was neutral on the question of whether the system should be symbolic or sub-symbolic. Building on this model, we might include not just sensory inputs, but add motor outputs. And drive action.

Eric Horvitz, Microsoft Research
In Pursuit of Artificial Intelligence: Reflections on Challenges and Trajectories

We are pursuing intelligence admidst inescapable incompleteness: limitations in representation, time, and memory. Uncertainty is ubiquitous. So machinery for handling uncertainty and resource limitations is foundational in intelligence.

Beauty and the Bottleneck
A lot of AI is about looking at the bottlenecks and gaining insights. This forces us to look at the economics of computation and bounded optimality. Why should we compute? Why perceive and reason?

Continual Computation
How can the agent make best use of idle cycles to plan and reason about the future? We can generate and use algorithms that model the tradeoffs an agent makes between the expected value of a computation, the expected value of information, and the expected value of exploration and learning. It boils down to looking vs. thinking -- how does an agent make these computational trade-offs? Much remains to be done here in AI.

Intention Machines (e.g. web search)
Taking query inputs and making inferences based on those inputs about information goals. This can be used in travel interfaces, too.

AskMSR: a heuristic search system that takes a query and predicts the likelihood of a correct answer based on material already on the web.

Mixed initiative interactions: The machine and human make independent contributions to solve a problem. The human does what it can do, the machine does what it can do, and each queries the other in achieving the solution. These systems are integrative because they span multiple systems (within the context of HCI interfaces). E.g. MS Lookout.

AI@50: The Future of Learning and Search

[AI@50]
Filene Auditorium, Moore Hall
Dartmouth College
Hanover, NH

Conference notes by Meg Houston Maker

Oliver Selfridge, MIT
Learning and Education for Software: new Approaches in Machine Learning

Selfridge will address learning: learning many things at once, and the full concept of learning in the human experience.

We know there is learning in other creatures. For example Alex, a 20-year-old African Gray parrot, has learned to speak English, and it can count up to six, and can name how many, e.g, small round objects are set before him. But this parrot uses a different kind of language than what Selfridge is really talking about.

And, Selfridge says he's also not going to talk about formalisms. There are enough people around here who will focus on formalisms [laughter]. Nor will he talk about neural networks.

Rather, he's focused on how we learn many things at once. We do learn things in order, but we don't have to learn them in order; sometimes we learn in parallel. For example, take a 3-month baby in a crib, with a sparkly mobile over him. He knows about reaching his arm, and about minimizing the visual angle so he can reach the object. And the first time he reaches he knocks the mobile away. But the next week, he may reach more gently. He knows to slow down a bit so he can control his movements. And much later, when the baby learns to kick a ball, he knows to slow down a bit before he hits it. It's a technique he's learned in other contexts. Selfridge says he doesn't care about how the baby does this formally. That's just not important.

But you can't apply a lesson learned about kicking a ball to how to eat spaghetti. There are complexities in learning. We don't read, e.g., just by recognizing characters. We look at a word or a phrase, and we connect it with meaning. We also look at a jar of mustard and we see Heinz, because we've connected past learning or understanding to that concept.

So, why do we learn at all? "To use technical terms, why do we give a shit?" The key role in learning is played by purposes. What are we trying to do? Goals and purposes play a large role in learning: if you don't learn, it's probably because you don't care.

And you never have just one purpose. You have a hierarchy of purposes. Every action or thought is a response to a purpose, something you're trying to do. When you learn chess, you don't learn just so you can win. You learn sub-structures that serve sub-purposes in the eventual service of winning. So when you learn the thing you're trying to learn, you learn lots of other things along the way. And this other kind of learning is curcial to the kind of responsive, rich behaviors we all have.

If you have a machine that plays chess, what use is that machine in teaching a child to play chess? When you teach a child to play chess, you lose so that the child will learn to take advantage of opportunities presented to her. This is a large part of ingenious education; to teach children to think for themselves and open the doors to new approaches to solve problems.

The role and importance of purposes cannot be understated. From the appreciation of beauty, all the way down to the movement of our muscles, can be expressed as control loops of purposes. Purposes get changed; circumstances arise. We have to fit the context around the purposes, and the purposes inside the context. Our notion of beauty when we are 8 is different from when we are 18.

And how do we get at the features that matter? There are relevant features that can't be taught, but which can be discovered. Information theory, which looks at channel capacity, was introduced in the 1930s. Now, young children know about it, and it's a feature that we know and use. The concept of the number zero is similar -- the feature of zero being part of a number was at one time a new concept.

How do we get new features? They can't just be generalized. So how do we learn new ones? Once you've gotten a feature, you can talk about it and teach about it. But if you don't have it, how do you get it? For example, how did Helen Keller learn to read? How do we get new concepts? What are the alternative ways of learning?

One thing about features is that if you want to make classifications and see what is similar between events, you need a set of features to describe them. How do we even learn what counts as a feature?

Be sensitive to the role of change. Life is always changing. And that change changes your point of view and your purposes. Part of learning has to be to understand that and to keep going anyhow. Marvin Minsky put it well when he said "the best is the enemy of the good." Meaning the best implies an optimization, which, when used in machine learning, implies that the world is standing still. Which is not a safe premise.

So, how do we express as a language for software not just what it should do but what it should try to do? At every stage the software should be told to try something and then to try something new. So instead of writing every detail of what we're trying to do, we should try to have the program work out what it should do. An important part of learning is that we don't learn nearly as much individually as we learn culturally. Let's put together a language to express many things we can't express now. Tell an AI program that it doesn't just want to do something, but to create a satisfaction parameter that will make it happier that it has done what it's tried to do. And remember, it's not just trying one thing, it's trying many things.

Ray Solomonoff, London
Machine Learning - Past and Future

Solomonoff is not interested in simulating humans. He's interested in creating machines that can solve problems much better than humans can. Like cure cancer. And these machines should follow Moore's Law.

At the beginning of AI, Solomonoff wasn't really impressed with the expert systems project. When you ask a human expert how they do what they do, they can't tell you. And these researchers had found that the more info you put into a system, the better it worked. This was contrary to most of science, in which when you get more info, you then can (or have to) reduce this to a set of rules or principles. This just wasn't happening in AI with expert systems.

Neural networks also seemed like an inefficient way to solve a problem. Given enough nodes, you can pretty much solve any problem, but it's cumbersome. Many other research projects are promising for some aspects of AI. But his approach was a different. He developed an algorithmic probability concept. Basically, he showed that you could take a universal machine (a Turning machine works, but others do, too) and give it random data, the output is infinite but can be predicted.

Solomonoff's math eludes me, but you can learn a bit more from his Wikipedia entry and his own website.

Leslie Pack Kaelbling, MIT
Learning to Think About the World

Early in AI there was a lot of emphasis on thinking and symbolic representation. Recently there has been emphasis on embodied cognition and embodied action; agents that have to perform in the world. A lot of robots can get by being reactive. But we need robots that can act in random environments, "in the wild," as it were. What is needed for that?

1) estimating the state of the world
2) making guesses about what will happen if you take any particular action

This means we need ways to represent uncertainty. The set of world states is too big to be useful. The problem is "learning to live." You do a lot of things, and you know a lot of things, so how can you figure out what action to take based on all of the things you do and know?

Probablistic dynamic outcomes: actions have probablistic outcomes, but the frame stays constant, and you jettison irrelevancies and minor nuances (MHM editorial: that seems mighty convenient). So now you have something that actually can be modeled.

She's talking so fast I can't keep pace with her! She shows examples from robotics, block worlds, and other systems. Check her website for more. I probably will have to.

Peter Norvig, Google
Web Search as a Product of and Catalyst for AI

Google's mission: Organize athe world's information and make it universally accessible and useful.

Machine learning and natural language processing are core technologies of searching now. Google offers a "predicted items" function that lets you do on-the-fly classification. E.g., if you type in Armani, you get Gucci, Prada, etc.

Trend history: maps the frequency of terms over time. E.g. of watermellon (spike in summer, low winter), and full moon (monthly), eclipse (predictably, when eclipses happen, plus a little noise for searches for the car and the programming language).

Natual language processing is a core technology. Here are some examples:

Spelling corection: E.g. "Britney Spears." Google takes a corpus-based approach. There is no "ground truth." Rather, they build a probablistic model based on the vast data of spelling variation in the corpus. The advantage of corpus approach is you will get names not in a dictionary, and you can acquire spellings in context, related to the surrounding words.

Extraction from semi-strutured text: it's possible to pull out facts from the text, too, like population data on a searched country: e.g., search on "what is population of japan" and you'll see how Google pulls out the population data and puts it at the top of search results.

Statistical Machine Translation (to, e.g., translate from Chinese or Arabic into English): There's no linguist on staff at Google who speaks these languages. They simply collect parallel texts (data that exist in both languages, e.g., from newspaper sites.), then they do a probablistic alignment; the results are extremely accurate.

Norvig takes a corpus-based approached, which is different from the traditional knowledge-base approach. Some concepts are not in reference books. The concept that "water flows downhill" is not something you can look up in the encyclopedia. Plus, even if you used existing knowledge, it would cost about $10,000 a page to encode, e.g., a chemistry textbook. So, how can we get knowledge or information without just paying someone to do it?

With Google, there is a huge corpus, much larger than what might be in an encyclopedia or textbook. And there is lots more accuracy with a larger training corpus, because you have more opportunity for disambiguation. This is "The Power of a Billion" (after Banko and Brill, 2001: Effect of Training Corpus Size).

Google created a complex model (called MapReduce) to process data. It's a learning algorithm that is broadly useful within the Googleplex and can be applied across projects.

Norvig references Andy Clark's book, "Being There: Putting Brain, Body, and World Together Again." We're not as smart as we think we are. We've figured out some clever stuff we can do with language and mathematics -- but these are external systems, or props, that amplify our intelligence.

Making information accessible and matching it to human needs is what Google's trying to do. "In the future search engines should be as ueful as HAL in 2001, but hopefully they won't kill people." (Sergey Brin)

Q&A:
Question: what about image processing?
Answer: we do very little image processing; Google Images is our second-most popular property, but the work is all around words surrounding the text. But we are starting to do a little image processing, signal processing, and the like.

AI@50: Third Poll

[AI@50]
59% say connectionism is a promising model, with some restrictions or qualifications

AI@50: The Future of Network Models

[AI@50]
Filene Auditorium, Moore Hall
Dartmouth College
Hanover, NH

Conference notes by Meg Houston Maker

Geoffrey Hinton and Simon Odinero, Toronto
From Pandemonium to Graphical Models and Back Again
Odinero gives this talk solo

How to build a perceptual system -- a "pandemonium architecture" after Selfridge

The good old fashioned neural network: input - hidden layer - output layer, plus a back-propagation algorithm as a derivitive of learning.

What's wrong with this model?

- you need a labeled training corpus
- it's not efficient; unless the weights are redundant, labels can't provide enough information
- the learning time doesn't scale well with more than 2 or 3 layers
- neurons need to transmit the signal forward and back-propagated information backward; real neurons don't do this

Overcoming These Limits -- using a Restricted Boltzmann Machine
A simple model is a 2-layer network (a shallow network): 1 layer of hidden units (with no connections between them) and 1 layer of visible units. Start with training data on the visible layer (images of reality), update the hidden layer, then update the visible units in parallel to get a "reconstruction" of reality. Then update the hidden units again. This network can learn to recognize features of the training data, but it tries to frame all perceptions in terms of that corpus (so if it were trained on the symbol 2, and was given a 3, it might recognize some features of 3 that correlate with those of the 2 (e.g., the top curve to the left).

It is also possible to create a more complex model that has multiple hidden layers. This seems more promising than back-propagation because there is only one type of signal sent through the system. This system can "fantasize" outputs that map to reality. So if you tell such a system that this symbol is a 2, it can create output shapes that look like 2. Likewise, if you give the system a hand-written 2, it can guess most of the time that it represents the numeral 2.

MHM editorial: There is no evidence that this system carries or expresses any understanding of the meaning of "two-ness."

Rick Granger, Dartmouth College
Essential Circuits of Cognition: The Brain's Basic Operations, Architecture, and Representations

Granger uses a Mac.

The Problem: Intelligence is undefined; it's only defined by example, by what we've learned. There's no formal spec. The only spec we have comes from brains. The question is how to scale from brain mechanisms to high level faculties. One approach is an analysis of brain circuits and systems, and a deriviation of alrorithms and data structures, not just statistics, and anatomically structured systems that construct high-level cognition. It's not a debate between statistics and algorithmic analyses. The issue is to show how logic arises from statistical methods.

As mammalian brain size increases, specializations decrease. In large brains, the posterior and anterior cortex are heavily connected, and the output of the anterior cortex, which in small mammals drives the musculature, has feedback mechanisms back into the anterior cortext. These feedback architectures are repeated redundantly in the brain in nested clusters and drives more complex behavior/output.

Granger shows an example diagram of a core loop within the brain (a thalamocortical circuit) that acts as a feed-forward network in recognizing, in his example, a flower. As the output of one region becomes input for another region, more complex patterns are created, generating a kind of grammar within the system. So, at the end of the day, there are a few algorithms that can be derived from brain circuits, and these are fodder for AI research.

For more, see BrainEngineering.org

AI@50: Second Poll

[AI@50]
71% of the AI@50 audience said statistical/probablistic methods are most accurate in representing how the brain works.

79% said an accurate model of thinking is impossible without further discoveries about brain functioning.

AI@50: The Future Model of Thinking

[AI@50]
Filene Auditorium, Moore Hall
Dartmouth College
Hanover, NH

Conference notes by Meg Houston Maker

Ron Brachman, Yahoo! Research
A Large Part of Human Thought

From the original summer project proposal -- the proposition that "a large part of human thought" was bound up in language, but there is also a conflation of language and concepts that language replaces:

"It may be speculated that a large part of human thought consists of manipulating words according to rules of reasoning and rules of conjecture. From this point of view, forming a generalization consists of admitting a new word and some rules whereby sentences containing it imply and are implied by others. This idea has never been very precisely formulated nor have examples been worked out." (From the original proposal.)

Cognitive Penetrability: language can produce widely ranging human behavior.

Knowledge Matters: "the city council refused to grant the protesters a permit becuase they feared violence" vs. "the city council refused to grant the protesters a permit because they advocated violence." In each case we understand a different reference for the pronoun "they." Background knowledge of the universe changes our interpretation of linguistic inputs, even if the inputs are, grammatically, nearly identical.

A good portion of human thought involves a store of represented beliefs, and procedures that operate on them to produce new beliefs (inferences) that impact decisions about what we do. The focus is on what can be known, and how, and what follows from what is known. Knowledge representation and procedures for manipulating knowledge became an object of study in early AI.

The internal conceptual monologue: From the simplest idea (formed from a simple syllogism such as "Socrates is a man and all men are mortal therefore Socrates is mortal") you get to daydreaming, imagining other possibilities, envisioning and caring about your new vision, planning about it, and teaching or preserving it as an idea. And all the way out, this action begets writing, reading, culture, and science. Having knowledge representations in our heads that are disconnected from concrete reality begets intelligence, culture, and society [MHM: some editorial liberty in this last bit].

You can use knowledge structures to do forward-looking inference, come to a mental conclusion, and make a decision about what to do. This was an original idea of the conference 50 years ago, and since then, it's had a great impact on:

- Database management systems
- Expert systems
- Descripton logics, and the semantic web (or OWL = web ontology language)
- Cognitive robotics
- Planning systems (e.g. SIPE)
- Model-based reasoning (e.g. at NASA)

Language and perception create usable memories, useful for making decisions about what to do now, or to plan future moves in the abstract. It's about taking advantage of past action, or learning. The explicit formal representation of what a machine knows and believes is at the heart of creating a machine that can do what humans can do.

To sum:
- Actions are conditioned by what we know and believe;
- At least part of this knowledge arrives in linguistic form;
- This allows us to focus on what needs to be known;
- And gives rise to a 'knowledge-based" approach.

This idea defined AI, as opposed to approaches like computational neuroscience.

Question: Much of the AI community has either forgotten this or would disagree with this focus. E.g. the natural language community.
Answer: These insights mentioned are still relevant, and essential to knit the whole of AI together.

Question: (actually a comment from John McCarthy): In some ways our proposal for a revolution was a proposal for a counter-revolution against the behaviorists that advocate a stimulus-response approach.

David Mumford, Brown University
What Is The Right Model for "Thought?"

Is thought logic/rule-based or is it stochastic models/statistical inference? This is an old question, but still being discussed. Logical deduction and statistical inference are really different, and maybe even incompatible models. Logic comes in many flavors: traditional, modal, temporal, etc. Stochastic models are differently structured. Logical deduction is brittle and unforgiving, whereas statistical inference is more adaptive and flexible.

"Logic/rule-based thinking is fun, but is it real? To grammarians, rules are the deep and very real way to understand natural language grammars, but language is a maze of exceptions within exceptions."

People think in a Bayesian (probablistic) fashion. People make inferences based on a collection of hypotheses about the context. E.g., given the following test:

Actual sound (missing consonant): "The ?eel is on the shoe."
Perceived sound "The heel is on the shoe."

People will not report they thought the word was wheel or steel. They know that shoes have heels, and so assume, absent other data, that the missing consonant was "h."

Implementing statistical inference: there are vast amounts of noisy data and many variables present. Statistical tables won't help you. So you must sample the entire probability distribution and do 'particle filtering' to create a result. But how does the brain do this? The brain isn't ever doing just one thing. There is no "grandmother cell" that fires when you think about your grandmother. The brain has to manage hundreds of variables at once, and filter. So choice is postponed.

Stuart Russell, UC Berkeley
The Approach of Modern AI

Russell would love to see the various branches of AI research merge; they have been quite divergent (cognitive science, mathematics, other research projects).

Formalisms or algorithms for probablistic knowledge bases: focus on concrete syntax, semantics, and completeness. A key component of any formalism is expressiveness. E.g. the rules of chess can be stated in about a page of first-order logic, whereas it takes 100,000 pages to do so in propositional logic. Humans operate on the 1-page version. The question is how we get there.

The focus of research projects should involve issues of various time scales, extended deliberation, a varied environment. A human-scale, dexterous 4-legged, seeing robot would be exceptionally valuable to the project.

AI@50: First Poll

[AI@50]
We use our polling dongles to vote on several questions. The most interesting question: When will computers be able to simulate every aspect of human intelligence? 41% of us said "More than 50 years," and 41% said "Never."