Saturday, January 28, 2006

Legends in AI: Chris McKinstry

From: Chris McKinstry
To: fuzzgun
Sent: Sat 27/10/2001 01:19
Subject: Software

I really like the stuff you're doing.

How would you like to become chief software developer for mindpixel?

Can't pay anything right now... but eventually...


I was very saddened indeed to learn of the death of Chris McKinstry. A particularly tragic and seemingly pointless way to go for an otherwise intelligent man. Back in 2001 I turned his Mindpixel job offer down, partly because he wasn't paying but also because I was quite happy working for my previous company doing industrial automation (which lasted up until 2005). Since then I've been a regular reader and commentator on his blog.

Mindpixel is one of the two main artificial intelligence projects designed to collect large amounts of commonsense knowledge from the general public via the internet. In his own words:
My primary inspiration for the project comes from observation: I observed that computers are stupid and know nothing of human existence. I concluded a very long time ago that either we had to write a "magic" program that was able to go out in the world and learn like a human child, or we just had to sit down and type in ALL the data
This is basically the same as the conclusion which Turing comes to in his 1950 article Computing Machinery and Intelligence. People had tried to create corpora of knowledge before, the most famous instance being Doug Lenat's Cyc, but by the late 1990s widespread internet availability made it possible for the same sort of information to be entered by anyone at any location. Using the internet is a much faster and more efficient way of grabbing large amounts of data from many volunteers, rather than just telling a couple of hapless university undergrads to type maybe a few hundred facts into a program.

Chris' aims with Mindpixel were quite specific, and motivated by the desire to build a computer program which could convincingly pass the Turing test. But this wasn't the kind of Turing test that you'll find in AI textbooks, or even in the original article by Turing himself. I think one of the things which Chris McKinstry will be most remembered for is his idea of a Minimum Intelligent Signal. This takes your bog standard Turing test and distils it down to the bandwidth of a single bit. Yes/no. True/false. Flip/flop. The simplest, stupidest, decision which any computer can make. Ask a person, or group of people, a set of yes/no type questions and you'll be able to gather probability data on what people believe to be true and what they believe to be false. If you do this with a large enough group of volunteers recruited from the internet you begin to be able to build a probability matrix for human psychology in general. If a computer program can give answers to questions with the same probability distribution then in this version of the Turing test you can say that it has passed. Collecting this sort of binary decision data also allows other questions to be answered. Does what people believe to be true or false depend upon the language they speak? Are there systematic relationships between individual beliefs? (so-called "world views"). By gathering large amounts of data and doing some careful analysis you can put some hard figures on what was previously just vague philosophical speculation (the common currency of AI discussion groups). Chris McKinstry was the first person to try to do that sort of analysis, forming a linguistic road map of the mind.
If the field of artificial intelligence is to evolve from hype and engineering into a full-fledged science, I believe measurement must become the core of the standard paradigm for research in the field. An automatically constructed corpus can never be a substitute for actual measurement in human subjects.
Roughly translated a robot or child machine which wanders about gathering knowledge is unlikely to acquire the same kind of mind as a human, and therefore would not pass anything resembling a classical Turing test. That's not to suggest that the robot wouldn't develop perfectly valid ways of thinking, just that it's thinking style would not necessarily resemble that of the average person due to all kinds of differences in physiology, social acceptance and so on. To build something which really walks and quacks like a duck you need to take measurements from the duck and incorporate that into your duck model.

As with virtually all AI projects, Mindpixel suffered setbacks and in 2005 - roughly corresponding with my own ISP woes - it lost its free server hosting. According to the Wikipedia entry over 1.4 million entries were made by the public, a small fraction of which were recently made available for research purposes within GAC-80K. Exactly what will be the fate of this data is now quite uncertain. According to his blog Chris was writing a new Mindpixel web site which would have allowed the data entry process to continue. This is now very unlikely to happen, and the long standing tradition within AI under these sort of circumstances is for all the data to be lost, archived where noone can access it, or sold to a company which then never uses it.

In my opinion since the data was entered by the public it should now become available to the public in full.

For the present the last man standing in this game is Push Singh with his OpenMind project. Singh's basic approach is similar to McKinstry's - leveraging the power of the internet - although the methods by which data is entered is somewhat more structured and not really aimed at mapping the psychological landscape. For an example of the sorts of practical uses which such knowledge databases can be put to a recent paper by Rakesh Gupta and Ken Hennacy is a good place to start.

Chris also had some other ideas which may have arisen from his work on Mindpixel. I didn't understand them, other than they were some form of mathematical quantum notion of brain function. He was particularly interested in one project of mine, where I was trying to automatically detect the layered structure of the cortex (the spatial arrangement of neurons - see the diagrams above) by analysing microscope images using image processing methods. He was also collaborating on a project with Cornell university designed to investigate the time and space related aspects of language.

Quite what Chris McKinstry's reasons for taking his own life were I've no idea. I was never very interested in his private life, and on the few occasions I emailed him the discussions were of a purely technical nature. Perhaps there were no reasons, or maybe he was suffering from some kind of mental illness or drug addiction.

I looked him up on Technorati, and was pretty horrified by some of the inept commentary. This one struck me as especially loathesome and ill-informed. There are a few others which are fair comment. So far I havn't seen any comments from anyone with any significant degree of technical understanding. Such minds it would seem are rare.

Towards the end of his article Mind as Space Chis McKinstry makes a clear prediction. Because of some fundamental mathematical properties he suggests that the idea that artificial intelligence will one day reach human level and then race far beyond it may not turn out to be the case:

And if mind is really founded on something mathematically like ψ, then there may actually be a fundamental limit on how much more powerful individual artificial minds can be in comparison to our minds. If all minds are constrained as ψ, then no matter their physical resources or lifespan, there is simply a maximum number of ideas that can be connected in a single entity that is related to the hyper-surface area of a hypersphere with a little more than seven dimensions. Like us, simulated people might do their grandest works cooperating in simulates cultures of limited individuals.
Of course there are plenty of people in AI who don't want to hear this, and it would indeed be a disappointing result if it turns out to be true. I guess only time will tell.

Also see References related to McKinstry's language hypothesis.


Arthur said...

Bob Mottram: You are to be commended for this kind, considerate article on the passing of a unique individual in independent AI scholarship. When I became active in the* newsgroups of Usenet about ten years ago, Chris McKinstry initiated a few exchanges of e-mail with me. He and I ("Mentifex") were somewhat in the same boat as AI outsiders. I mentioned his Mindpixel work here and there on my webpages, mainly as an example of how NOT to do AI. I wasn't always as nice to him as I could have been. Now I wish that I had been nicer. Requiescat in pace. -Arthur (Mentifex)

Bob Mottram said...

Well I know he was widely criticised and considred a marginal figure. Most people criticised him from a standpoint of pure ignorance though, saying things like "well, this guy is just collecting a bunch of yes/no questions and answers. Big deal". On a superficial level that's exactly what he was doing, but considering it only at that level would be missing the point entirely.

Unlike a lot of the more mainstream AI researchers Chris had a few very specific hypotheses and a proposed method by which they could be tested and validated or falsified in the true spirit of science. His minimal MIST test is useful because it allows measurements of degrees of humanness, whereas a traditional Turing test just makes a boolean decision - person or machine.

The line of research he was persuing I think remains a valid and unanswered one. It will probably become known as the McKinstry hypergeometric hypothesis. He gave us the idea and a method of testing it, so if anyone wants to pick up the ball and run with it the challenge is still open.

Arthur said...

Mitchell Porter has posted a description of the McKinstry hypergeometric hypothesis -- perhaps something for The McKinstry Institute to work on.

Carlos said...

Bob, I knew Chris for 6+ years. Personally and you may called professionaly. We set on his appartment for long talks at night about science and computers, but also about just funny stories.
He was a troubled person, with some inner shadows he couldn't fight back for good. But also a funny, smart and good-hearted man.
I did some coding for on the old server. I have "touched" the database. But recently Chris was moving it from one server to another. I have spend this lasts days since the tragedy on getting his stuff, papers, reseaarch and also DB together. I' have family contact i hope will lead to further collaboration on this issue.

Bob Mottram said...

He was writing a new paper on his geometric hypothesis - A General Geometry for Cognition and Consciousness. I don't know whether this ever got beyond being an abstract, or whether it was due to be published.

If the Mindpixel database can be salvaged I think it should be. If there's still sufficient interest perhaps the project could be continued.

amethyst42 said...

I dated Chris in High School for a short while. Even then he was emotionally unstable. He was brilliant, but had bizarre ideas that affected his credibility. I don't know if he was ever diagnosed as schizophrenic, but he fit the profile well.

Tim Tyler said...

The size of the large data centers that are springing up around the world strongly suggest that any limit is far off. Of course, there's the Bekenstein bound...