Chris McKinstry Replies: Telecopes, AI And More

1) GAC
by Dungeon Dweller

I have an active interest in artificial intelligence. I study it as part of my major, and hope to do research in it in the future. As a young man coming up in the world, I am hoping to enter into research eventually, am entering into research at my university (WVU).

Your project reminds me of several projects/theories that have been discussed before. In the psychological debate, your system depends entirely upon nurture, it would seem. I like that kind of system and research. I do have a few questions.

1. What separates this from other projects in the field?
2. Where did you draw your inspiration for this project?
3. What kind of support staff do you recommend to an individual who has never led research before, but would like to? (I ask this of many of my professors who conduct research)
4. Where are you getting the bulk of your input for this project?
5. What do you hope to learn from this project?
6. At what time will you consider this project a success?

Chris McKinstry:

Question 1-1:

There are three primary features of the MindPixel Digital Mind Modeling Project (also known by GAC -- for Generic Artificial Consciousness -- which is public interface to the project) that distinguish it from other large scale knowledge projects such as CYC.

1. The first phase is a completely public, internet based effort. All the data it will be collecting will come from average people, with no specific training in AI or psychology. It is like seti@home in many respects, except that we're not after your CPU's cycles, but rather your humanness. We're actually seeking to extract the entire content of an average person's mind bit by literal bit from millions of different internet users. We're not trying to write the algorithm for consciousness, but rather create the world's most rigorous fitness test (a Dawkinsian continuous variable) and get it into the hands of researchers who will attempt to make systems that will learn or evolve into consciousness by feeding back against this fitness test. Not only will we be collecting consensus fact, but also consensus emotion. (When the project is fully operational, in addition to collecting information about each MindPixel's truth or falsity, we will also collect emotional data based on Mehrabian's PAD model of emotion.)

2. The second phase of the project involves releasing the data collected to the scientific community and providing those researchers with some funds (generated by advertising to the people supplying the data) to conduct their research. As a side note, Jeff Elman's page contrains information about recurrent neural networks that are very good at processing just the kind of data that this project will collect and distribute. Specifically his 1990 article, Finding Structure in Time (PDF) is one of the most important neural network papers ever written; it strongly influenced me.

3. Finally, the project is a meritocracy. People will gain voting rights that will give them a say in every aspect of how the project is run, from data collection and use to the distribution of data and research funds, based entirely on the amount of data they have contributed to the project. The more work you do, the stronger your voice becomes.

Question 1-2:

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. When I was studying psychology in the late 80's I wanted to begin to gnaw the bullet and start getting people to type in ALL the data. It was my plan then to get people to enter data as part of an intro psych course, or get the university to allow me ask people for data when they logged on to the university's computer system. I was never able to get permission for either and the idea sat on the shelf until I downloaded my first copy of NCSA's Mosaic in 1994. I saw in following my first hyperlink, a different path.

I decided to collect my data via the internet. But, the problem was, that I needed to think of a standard format for the data; some way of representing human knowledge that an average person could learn quickly. That idea didn't come to me until I was preparing an entry for the 1995 Loebner Prize. Jackie, my program, was a stimulus response creature. You would ask her a full text question, and she scan her database for a canned full text response. My idea for the Loebner competition was to have her talk to a lot of people a get a lot of canned responses (at the time I was consulting for a large insurance company and brought Jackie to work everyday where she could talk to my colleagues) As well, I stored the responses in a number of different ways: phonetically using soundex, again with all the words in each stimulus sorted alphabetically, and also with a primitive concept token system. So, if there was no direct match, she would look for a phonetic match or sorted or conceptual match. Essentially I was breaking down each stimulus and standardizing it like a Fourier transform breaks down a waveform.

Then suddenly Hugh Loebner changed the rules. No longer was passing a text based Turing Test good enough for him. Now he would only award his prize if the system passed a full audio/video Inquisition. I hit the roof! Hell, there were tens of thousands of people on the planet that couldn't pass that kind of test! Anyone blind or deaf are just two obvious examples. I withdrew Jackie in a loud protest, stating that intelligence didn't depend on the bandwidth of the communication channel; intelligence could be communicated with one bit! If you locked a person in a box I could detect them with a series of yes/no questions and nothing more. And there all of a sudden, I had my answer (and a quick paper - The Minimum Intelligent Signal Test - An Objective Turing Test in Canadian Artificial Intelligence, issue 41.) There was a minimum intelligent signal, and it was just one bit. I would store my model of the human mind in binary propositions. I would make a digital model of the mind.

I realized within minutes that a giant database of these propositions could be used to train a neural net to mimic a conscious, thinking, feeling human being! I thought, maybe I'm missing something obvious. So, I emailed Marvin Minsky and asked him if he thought it would be possible to train a neural network into something resembling human using a database of binary propositions. He replied quickly saying "Yes, it is possible, but the training corpus would have to be enormous." The moment I finished reading that email, I knew I would spend the rest of my life building and validating the most enormous corpus I could.

Question 1-3:

Support staff! I recommend using the entire planet as support staff! Seriously, don't even dream about it. Almost every researcher I know works on their own or with a handful of collaborators. When you're a big cheese you might get a student or two, but other than that you'll get nothing more than shared use of a departmental secretary. You'll definitely be writing all your own code for a very long time.

Question 1-4:

I can't tell you that yet because at the time I wrote this, the project was not yet online (should be now though.) What I can tell you is that in 1995 I did try to collect this same data, using a web based form that sent an email back to me. I managed to collect some 450,000 items. This time, I expect to collect more and higher quality data and I expect it to come from a wide cross section of the internet public. I should also note MindPixels will be collected in multiple languages, which opens up the future prospect of mapping the sampled human languages to each other concept by concept. It will be very interesting to see exactly how an artificial consciousness trained in English differs at the conceptual level from one trained in say, Spanish.

Question 1-5:

I hope to learn what the human conceptual network looks like. I hope that in a few years I will be able to access a map of all the concepts in the head of an average person or to have learned why I can't.

Question 1-6:

I will consider the project a complete success when the cover of Science announces that for the first time in history there exists an artificial system that has passed a scientifically strong form of the Turing Test known as the Minimum Intelligent Signal Test.

The slightest movement ought to mean millions of miles so thoes pesky little earthquakes should be a problem. Not to mention how you guys move the telescope accurately?

Chris: You're quite right about the system being very sensitive; if I walk on the azimuth platform of a VLT telescope while we're observing, I will destroy the observation. For normal tracking we use a software system called Tpoint written by a well known telescope genius named Pat Wallace (Pat has a wonderful and detailed article about telescope pointing that anyone seriously interested in telescope pointing should read); the same system is in use on telescopes all over the world. Basically what we do is build a pointing model for each of our telescopes. This involves pointing each telescope to a number of different points uniformly covering the sky. At each sample point, we observe a guide star and record how it moves from the center of the field over about one minute of tracking time. After we have collected enough data, we build a computer model of the telescope's tracking error. Then we basically run the model backwards into the telescope control system and thus apply corrections that try to cancel out the tracking errors of the telescope. This of course can't correct for any unusual vibrations, we rely on normal guide star tracking and hydraulic isolation of the telescope for that. And baring a large earthquake, Tpoint, automatic guide star corrections and the isolation work pretty well (In the event of a large earthquake, there are giant airbags that inflate to protect the mirror from damage.)

As a simple example, one could compute the differences between a sequence of pictures of the same portion of the sky, looking for anomalies like giant asterioids on their way to wiping us all out.

Chris: seti@home is one of the most impressive demonstrations of how the world of science has changed. There are now over 2 million average people working together for a common scientific goal. I just wish they sold advertising to raise funds for other worthy (meritocratically determined) projects. It really bugs me that my Pentium III 450 which has done over 7000 hours of seti@home processing since last June, hasn't shown me a single science supporting ad. What a waste!

Now as for your idea of doing the same thing in optical wavelengths, I think in it there is a great deal of merit. There are a whole pile on new survey telescopes coming online soon that will be useful for just what you proposed. And if you read ahead to my answer to question 11, you'll see I do think it is a problem we have to pay attention to. (As well, I know of at least two virtual telescope projects; the NRC's National Virtual Telescope. See NVO White Paper (PDF) and ESO's ASTROVIRTEL which seek to allow data mining of previously collected telescope data.

In general, I think the future will see a lot more distributed processing projects doing useful science. The question remains whether or not it is more cost effective to build supercomputers for critical projects or harness the CPU's of private citizens, and I think the answer will need to be determined on a case by case basis. As well, there will be some projects (my own for example) where the CPU cycles are incidental; where what we want to harvest is not your electricity and capital equipment, but actually your humanity.

I am continuously frustrated that people's general perception seems to be that science and art, spirituality, and so forth are divided by an uncrossable schism. People feel the need to pit science against spirituality; logic against intuition. It is a rare thing that people accept the idea that these are different ways of approaching the same reality. As a dreamer and artist as well as a respected scientist, what do you say to people who doubt that scientists can be spiritual and artistic people?

Chris: Science for me at least, is concerned with the external, the measurable; while art is concerned with the internal and immeasurable. Every scientist knows measurement can only go so far; that nature at its most fundamental is immeasurable. Unfortunately many scientists turn away from what they can't measure (and conversely, many artists turn from measurement) instead of finding some way, any way to express it. It is this turning away or fear of the immeasurable (or many artist's converse fear of reduction to measurement) that creates doubt; that separates science from art. The scientist can learn that one does not become any less of a scientist for attempting to express the inexpressible or attempting to measure the immeasurable, just as the artist can learn that because we are neurons and our neurons atoms, doesn't mean we are any less human.

But what do you do with them?

How parallelizable is the problem of micro-adjusting small portions of a large deformable mirror to correct for atmospheric distortion?

I remember a Scientific American article stating that you'd have to devote a top-of-the-line Cray to continuously recalculate the deformations needed given data from the guide star, or laser simulated guide star. If this problem is highly parallelizable, you may be able to get away with _much_ cheaper hardware.

I could have been mis-reading the article, but it seemed to me as though the idea was to build a single-mirror system. On the other hand, in radio astronomy, and in the insect world, arrays are considered the norm. Is there some advantage that a single mirror gives that cannot be duplicated using multiple smaller mirrors? (Simpler optics is an obvious one, paradoxically. :) Or is this (at least in part) NerdTrek III: The Search for Sponsors, where a record-setting single telescope is going to get more interest than a comparable array?

(A supplementary question, to go along with this. Let's say, for the sake of argument, that optical arrays are practical. Do you see any possibility of optical astronomers adopting the same line as radio astronomers, in trying to build an effective 1Km+ optical telescope, using an array?)

What are the benefits of having an Earth-bound, optical telescope? Or rather, what can a larger optical telescope find better from Earth that we can't already find on other wavelengths and from other venues (i.e. The Hubble)?

I noticed in your 'fave books' section that you have the blind watchmaker, et al.