Coverage of the 25th Anniversary of AAAI at the 20th National Artificial Intelligence Conference

This kind of thing only merits local coverage, but interesting to read about the robot game playing, etc. Excerpt below (full story)

More than 1,000 researchers, technologists and analysts are gathering here
starting today for the 20th National Artificial Intelligence Conference to learn
about the latest trends in AI science and technology.

Another highlight will be the conference's first general game-playing competition. Unlike specialized game-playing systems designed especially for chess or tick-tack-toe, for example, general game players are computer systems able to accept formal descriptions of arbitrary games and then play those games without human intervention. The winning entrant will receive a $10,000 prize.
Robots aren't serving us breakfast in our homes yet, but artificial intelligence has emerged as a quiet, but significant force in our lives, said Tom Mitchell, artificial intelligence expert and director of the Center for Automated Learning at Discovery at Carnegie Mellon University.
"We still aren't at the point of having systems that can behave as intelligently as you and me," Mitchell said. "But we're a lot further along in terms of building computers to do useful tasks in perception, natural language processing, and planning and scheduling."
For example, every time your credit card gets swiped, machine learning -- a form of artificial intelligence -- is used to decide whether to grant you approval. Internet search engines, the systems used to read the mail at the post office and luxury cars with "smart cruise control" all rely on AI to work.

Captology

Captology New concept for this week is captology. Captology is the study of the intersection of computers and persuasion. You can read more about captology on the home page of the Stanford University Persuasive Technology Lab: captology.stanford.edu. Of course, there's a Wikipedia captology entry as well.

There’s a whole science of persuasion, best captured, apparently in the works of Robert Cialdini, here is an excerpt from Wikipedia:
He is best known for his popular book on persuasion and marketing,Influence: the Psychology of Persuasion (ISBN 0688128165). His book has also been published as a textbook under the title Influence: Science and Practice (ISBN 0321011473). In writing the book, he spent three years going "undercover" applying for jobs and training at used car dealerships, fund-raising organizations, telemarketing firms and the like, observing real-life situations of persuasion. The book also reviews many of the most important theories and experiments in social psychology.
Captology books lead elsewhere

Robert Cialdini’s book on persuasion (referenced in the previous post) seems like a great place to start – rather than diving right into the more specialized field of captology. Even so, a little poking around on Amazon and on the Stanford site reveals some likely good books on the subject, and sends me on the trail of some other books that seem promising – including a surprise.

The Persuasion Handbook: Developments in Theory and Practice by James Price Dillard

At $135, this looks like a library check-out rather than a speculative purchase. One of those books cobbled together from journal articles. These can be a good way to get up to speed on the current research in a field, but I may not have the background to grasp it all.


A fairly recent book, that “explores persuasion by considering its antithesis: resistance..” Unfortunately, the only recommendation is by another author, possibly trolling for free advertising. But worth looking at I think, also appears grounded in scholarship – and cross-linked by Amazon with the previous book.

Persuasive Technology: Using Computers to Change What We Think and Do (Morgan Kaufmann Series in Interactive Technologies) by B.J. Fogg

The reviews for this one are quite mixed, some saying it is banal and creates needless taxonomies, others saying that everyone should read it. It’s out of the Stanford lab, though, so I think it’s worth a look.

The most amazing thing, though, is that an Amazon recommendation for this item leads to:
Emotional Design: Why We Love (Or Hate) Everyday Things by Donald A. Norman

I can’t believe I haven’t already read this and don’t own it. The Design of Everyday Things (the “teapot” book) is one of the most influential books about computers that I’ve ever read, although really it’s about the design of anything that people interact with.

This led to a book more specifically on computer design:
Paper Prototyping: The Fast and Easy Way to Design and Refine User Interfaces (Morgan Kaufmann Series in Interactive Technologies) by Carolyn Snyder

The concept here is simple: prototype your app by holding sheets up paper up with the interface crudely drawn on it, ask them to use it.

And finally, a book about an area anyone (especially me) can definitely use some help.

Social Cognition: Making Sense of People by Ziva Kunda

Fighting disease with machine learning and statistics

The previous post about the potential for treating a subset of patients differently based on data mined from patient information and outcomes raises the issue of the paucity of such data today. Extracting patterns from large data sets is one of the things that machine learning is really good at -- if only we could get at the data needed to draw sophisticated conclusions from the silos it is currently hidden away in (or from the ether into which it vanishes, unrecorded). Truly remarkable things would result in terms of treatment if reliable statistics about outcomes, treatment methods, symptoms and best practices were kept. Even more could be done someday when DNA records are attached (as was used in the breast cancer research). Although little attention is paid, many are aware that mistakes could be avoided using these kinds of statistics, but there is even less awareness of the potential new treatment options that would arise -- as in the breast cancer study, where high risk patients with cancers much more likely to metastasize could be treated differently. Making this a national priority could have more impact on health than any comparable undertaking -- as evidenced by the incredible advances made in anesthesiology when accurate statistics were tracked. Those advances were primarily aimed at mistakes, not breakthroughs in treatment and, if nothing else, keeping these statistics would allow us to fairly evaluate hospitals, doctors and their differing techniques -- current raw metrics, like death rates, as is often noted, ignore the severity of cases. Machine learning techniques could take all recorded factors into consideration to estimate the impact of choosing a particular hospital or doctor. Today information about the benefits of small differences, such as the provision (or not) of hand sanitizers requires considerable investment (can't find the reference right now), but could be done statistically if data was routinely collected. For more information about mistakes in medicine, including the impact when anesthesiologists decided to put patient's health ahead of their own desire to avoid facing their mistakes, I suggest the entertaining and sometimes dramatic book Complications, by Atul Gawande

Potential for more personalized medical treatment illustrated by breast cancer study

Using machine learning techniques, researchers have found a subset of breast cancer patients that have much worse outcomes than is typical -- "the 10-year metastasis-free probability being only 24% for the poor group, compared with 85% for the good group." This is pretty remarkable, although the confidence interval is pretty huge:

Here, we show that within a subset of patients characterized by relatively high estrogen receptor expression for their age, the occurrence of metastases is strongly predicted by a homogeneous gene expression pattern almost entirely
consisting of cell cycle genes (5-year odds ratio of metastasis, 24.0; 95% confidence interval, 6.0-95.5)

The key thing here is that this is a subset of a subset -- the kind of pattern difficult to extract without statistical methods. The authors go on to say:

The methods described here also illustrate the value of combining clinical variables, biological insight, and machine-learning to dissect biological complexity. ... Our work presented here may contribute a crucial step towards rational design of personalized treatment."

The study was publised in Cancer Research (A cell proliferation signature is a marker of extremely poor outcome in a subpopulation of breast cancer patients. Cancer Res, 2005;65(10):4059-4066).

When will computer hardware match the human brain? by Hans Moravec

Blue Brain Project attempts to simulate human brain

Using 1 processor per neuron, the folks at IBM, in partnership with the EPFL (a college in Switzerland) intend to simulate a brain component, the neocortical column "a cylindrical element about a third of a millimetre in diameter and three millimetres long, containing some 10,000 nerve cells. It is these columns, arranged side by side like the cells of a honeycomb, which make up the famous “grey matter” that has become a shorthand for human intelligence."

Should Moore's Law continue resolutely up the power law curve, Charles Peck, the leader of IBM's side of the collaboration, reckons it should be feasible to emulate an entire human brain in silico this way in ten to 15 years. (I guess I hadn't realized we were that close.) Of course, they mean that IBM will be able to afford to construct one, not that we will all have one on our desktop.

This should make Jeff Hawkins, author of On Intelligence, happy, since he posits in his book (if I recall) that these structures provide the recall-based, predictive source of our intelligence.

The article says that the most interesting questions "surely, are whether such an artificial brain will be intelligent, or conscious, or both." I guess I am more short-sighted -- the questions that interest me the most are ones like how you will boot such a brain -- will it be self-organizing? I don't think that our brain is entirely so -- surely most have accepted the idea that we aren't a blank slate. And even if and when we either correctly mimic the existing structures or design functional alternatives, other problems will arise. We know that real brains need constant stimulation to develop correctly -- how will they provide that to a massive computer locked in a building? After all, a child raised in a lab, without play or true "human" interaction would be severely disadvantaged if not retarded. (Perhaps a mobile "cute" point of presense to interact with the real world?) Should the inputs be imitations of ours -- artificial eyes/ears/touch? Or entirely novel inputs, like the stream of new pages posted to the Internet with interactions through chat rooms, online games and blogs?

Of course, lots of other questions come up -- can we still experiment on these brains when they get smart enough? What will this kind of power do to the field of AI and machine learning? I'm sure the science fiction authors are working on the solutions to some of these...

Artificial Intelligence Meets Ancient Ritual in New Wedding Planning Software

This is really not what I'd call "artificial intelligence" and it doesn't involve any machine learning, but it certainly is somewhat amusing. (Warning: the article is a press release, so take mental precations!) This program uses what you might call an expert system to suggest the wording for your wedding invitations and to do things like expand abbreviations (you aren't supposed to use abbreviations on your formal invitations, according to the article). So, for example,

... the software figured out that a street address ending in “LA CT” means “Louisiana Court”, but a street address ending in “CT LA” means “Connecticut Lane”

This is all fun to read about, but is fundamentally a downer for me -- it just serves to remind me how far we have to go before we have truly functional intelligent systems that can reason about the activities we are doing and help us do them. Just thinking about the work left to do gives me the willies.

Imagine trying to shoehorn all of this kind of "intelligent" technologies into today's infrastructure of API's and libraries -- Windows could add an abbreviation expander API, but if all APIs of similar functionality were added you'd need an expert system just to help you put all the right APIs together... But wait, you already need an expert system in order to figure out the current APIs! (or to become an expert yourself, which amounts to the same thing, right?)

Link: http://www.emediawire.com/releases/2004/9/emw160666.htm

The Long Tail

Forget squeezing millions from a few megahits at the top of the charts. The future of entertainment is in the millions of niche markets at the shallow end of the bitstream.

When working on collaborative filtering systems, we’ve long wondered how much of the action was in the low volume items in the “tail” of the Zipf distribution that these things always seem to fall into. For some items, such as movie-theater movies, there is almost no tail, since low-popularity movies just don’t get made or distributed. For others, such as books, it’s clear that there are a lot of members of the tail, but we never knew how much mass was in there. An article in the latest Wired magazine (I waited until now to share it, since they waited a while before putting it online) posits that the tail for many items is big and profitable, and that recommender systems (including collaborative filtering systems and human recommendations) are pushing more and more traffic to the broad base of less-popular items. Consider this:

The average Barnes & Noble carries 130,000 titles. Yet more than half of Amazon's book sales come from outside its top 130,000 titles. Consider the implication: If the Amazon statistics are any guide, the market for books that are not even sold in the average bookstore is larger than the market for those that are (see "Anatomy of the Long Tail").

In addition, many times these unpopular items are more profitable, for example when the customer buys a music track from an album from decades ago. There are a lot more interesting tidbits in the article, and some implications for automatic recommender system design, since it’s tougher to make good recommendations about the more sparse data, which may also need different parameter settings for best performance then those that give best results for the very popular items. It gets even more interesting should it turn out that you want to bias your results towards the less popular items. One more conclusion: this seems to indicate there is even more of a value in being the highest-traffic site (as in ebay and amazon) – only if you have enough traffic in the tail can you make quality recommendations about the data in the tail, driving more traffic and profits, enabling you to be more competitive on price…

Link: http://www.wired.com/wired/archive/12.10/tail.html

Oh, there also was a Slashdot discussion about it yesterday, with typical low S/N ratio: http://slashdot.org/article.pl?sid=04/10/05/185236&tid=188&tid=187

Support Vector Machine used to ID cars vs. scenery

Spanish researchers have built a cruise control system that tries to intelligently avoid other cars on the road. The system distinguishes vehicles from other objects using a SVM (although the article refers to it as "support vector matching," I'm assuming it's the SVM we all know and love.). I notice that the article doesn't really say how well it works:

Sotelo admits it will be crucial to improve the overall reliability of the system in order to persuade people it would work safely. Another key problem is that the system has trouble working in low visibility.

The visibility problem can be solved by adding radar, but it would be nice to have some info on how well it works under even optimal conditions.

Link: http://www.newscientist.com/news/news.jsp?id=ns99996471

Book Release: AI Game Engine Programming

AI means something slightly different to game programmers than to computer science researchers. In general, it seems like game programmers have a broader, less sophisticated, but more practical approach to AI programming -- anything that you need to know to program an AI, either a sophisticated opponent in a strategy game or a NPC in a first person shooter, or even a realistic car driving down the road. I plan to read some books in this area soon, since I'm just generally interested in this topic. This link isn't a recommendation for this particular book -- it's just a new release in this area. Sometime soon, I'll post a survey of the books out in this area and pick a few to actually read and review.

This book in particular tries to be a general survey, here's a bit from the publisher:

AI Game Engine Programming provides game developers with the tools and wisdom necessary to create modern game AI engines. It takes programmers from theory to actual game development, with usable code frameworks designed to go beyond merely detailing how a technique might be used. In addition, it surveys the capabilities of the different techniques used. In addition, it surveys the capabilities of the different techniques used in some current AI engines, and covers common pitfalls, design considerations, and optimizations. If you're having difficulty determining which techniques to use, or looking for working code best suited to a particular game, you'll find the answers here.

Link: http://www.amazon.com/exec/obidos/tg/detail/-/1584503440/qid=1097000700/sr=1-85/ref=sr_1_85/103-8352963-9919842?v=glance&s=books

If, for some reason, you don't like to visit Amazon, you can find the publisher's site (with the same blurb that's on Amazon) at http://www.charlesriver.com/titles/aigameengine.html

Artificial neural networks and DNA microarrays to successfully predict clinical outcomes of neuroblastoma patients

This is actually much more interesting than it sounds. Basically, these researchers used neural networks to train a model that could predict the patient outcome from the patient's genes:

First, the researchers performed gene expression analysis using cDNA microarrays containing over 25,000 genes to create global gene expression profiles of primary tumors from 49 patients diagnosed with NB whose clinical outcome was known. The patients were divided into either good (event-free survival for greater than 3 years) or poor (death due to disease) outcome groups. "Setting aside independent test samples, neural networks were trained to recognize or predict 'alive' or 'dead' expression profiles from the remaining samples," said Khan. "Then we determined if we could predict the outcome for the test samples using these trained ANNs." They found that the ANNs could predict the clinical outcome from any individual gene profile with an accuracy of about 88 percent.

As these gene profiles consisted of over 25,000 genes, the researchers tried to optimize the profiles and find the minimum number of genes that could act as a predictor set. The ANNs identified 19 genes whose expression levels could accurately predict clinical outcome. When only looking at these 19 genes, ANN prediction accuracy increased to 95 percent (!! ed)

Using the 19 predictor genes, the ANNs were also able to partition the subset of patients classified as high-risk into good and poor outcome groups. "What was most exciting," said Khan, "was that we were able to predict which of the high-risk patients would fail conventional therapy. This has major clinical implication since we are now able to distinguish a group of ultra-high-risk patients who will not respond to conventional therapy and therefore require alternative treatment strategies. We may also be able to reduce the intensity and thereby reduce the toxicity of treatment regime to those predicted to survive based on their gene expression profile."

"And since we are using 19 genes instead of 25,000," Khan added, "we can translate our findings to the clinic because simple prognostic assays can be developed based on this small number of genes. In fact, three of the genes found to be over-expressed in poor outcome tumors encode proteins secreted into the blood, meaning they could be used as serum prognosis markers in a simple blood test."

Wow -- that's pretty cool.

Full story: http://www.news-medical.net/?id=5222

Program Cracks Crosswords

A new program solves crosswords using similar techniques to those used by Eric Brill et al at Microsoft Research to answer questions:

It's a boon for puzzle addicts and a small leap forward for artificial intelligence: a computer program that can solve crosswords in any language.The program, called Web Crow, reads crossword clues, surfs the web for the answers and fits them into the puzzle. Computer engineers Marco Gori and Marco Ernandes at the University of Siena in Italy say a prototype should be available by the end of the year.

http://www.nature.com/news/2004/041004/full/041004-2.html