(August 2001) [Printed in "Reality Module No.23" as "Freeform Futurology V."]

FREEFORM FUTUROLOGY (6)
(A casual series of articles exploring various aspects of our evolving society)

Artificial Minds? (AI Revisited)

Related Articles

In 1996, world chess champion Gary Kasparov accepted the challenge of a computer, IBM's Deep Blue chess-playing program. Kasparov was shaken to the core. With 32 microprocessors, Deep Blue could analyze 200 million positions per second.

"I could feel - I could smell - a new kind of intelligence across the table," Kasparov admitted. "I got my first glimpse of artificial intelligence ... when in the first game of my match with Deep Blue, the computer nudged a pawn forward to a square where it could easily be captured." It dawned on Kasparov that for the first time he was facing a machine that could see ahead in novel ways. "I was stunned by this pawn sacrifice," he admitted.

In the first match, although Deep Blue took the first game in the series, eventually Kasparov found its Achilles' heel and trounced the computer, 4 to 2 ... Kasparov found the weak spot of the computer: chess-playing machines pursue a set strategy. If you force the computer to deviate from that strategy, it becomes helpless, flailing like an overturned turtle on its shell. "If it can't find a way to win material, attack the king or fulfil one of its other programmed priorities, the computer drifts planlessly and gets into trouble," Kasparov said. "So although I think I do see some signs of intelligence, it's a weird kind, an inefficient, inflexible kind...."
[Kaku, Michio. Visions: How Science Will Revolutionize the 21st Century and Beyond. Oxford University Press, 1998. pp.60-61. ISBN 0 19 850086 6]

I last touched on Artificial Intelligence in 1998 in my essay "Prophets of the Silicon God." [RM2] (This essay will also refer to elements of my "Is There Meaning In Dreams?" series of essays [RMs 9, 10 & 11].)

I have found Kaku's book to be very informative on the subject of Artificial Intelligence.

Firstly there are two approaches to AI - the 'bottom-up' school and the 'top-down' school. Both are relevant.

The bottom-up school works with neural nets and robots. The robots learn by trial-and-error how to move and how to interact with objects in the real world - much the same way that babies learn how to judge distances with their eyes & to clasp objects - and later how to balance and how to walk.

We end up with robotic insects crawling across the floor and avoiding objects - or maybe finding their way through mazes.

(A far cry from the time-lapse films of the 1970s of early robots moving a short distance, and then spending hours calculating their position before taking another step.)

The robots have become proficient at moving about, grabbing things, and generally interacting with the objects of the world.

The bottom-up robots are self-trained, learn as we do how to work in the world. (They do the things - like moving around - that we do unconsciously, without "thinking.")

The top-down school of AI is more traditional - it is rule-based programming (heuristics) and attempts to model reasoning. It is based on very complex decision-trees and is, in effect, an extension of the concept of the Expert System.¹

¹Expert Systems can use a sort of question-and-answer system, where you provide the answers - to do something like diagnose blood diseases.

Expert Systems work very well within their narrow specialities - but give them a problem which lies outside their area of expertise and they flounder, or give nonsense answers.²

²Like the Medical Expert System which diagnosed a rusty car as having measles.

An AI must be a generalised Expert System - and so researchers attempt to map out rules to cover all areas of human experience. (From the elementary - "If you are holding an object and let go, it will fall to the ground" - to the complex "If you are at a birthday party and you give the birthday person a gift which is identical to a gift which someone else has already given to them, you will have to take your gift back and exchange it for something else. This rule does not apply if your gift is money.")

We end up with hundreds of millions of lines of code - but still the AI makes elementary mistakes. (There are always rules so basic we never realised they had to be included.)

In short the top-down school is an attempt to model human reasoning - the so-called "higher brain functions."

The failure so far to produce a generalised AI is not a result of heuristics being a 'bad idea' - it is because developing a generalised AI is such a highly complex task.

What we have with these two schools of AI research are the modelling of two areas of human intelligence - our subconscious interacting-with- the-world and our higher level reasoning ability.

We don't teach children thousands and thousands of heuristics - we teach them basic rules, and they reason out more complex rules from simpler ones or learn through experimental trial-and-error. (A top-down machine needs to be able to learn. The machines are capable of combining existing rules to produce new rules and conclusions from these rules, but this may not be quite enough.)

In my original "Meaning in Dreams" essay I wrote about how each of us has a 'working model of the world' in our heads, and how we use this model in understanding the world and in making predictions about things such as how other people will behave.

Which school of AI is this 'working model' more analogous to?

Both - like the bottom-up school it is a neural net of associations which is learnt & modified, but like the top-down school it is based on rules and associations.

When the top-down and bottom-up schools meet we will be providing an Artificial Intelligence with a working-model of the world which can be interacted with and have inferences made from.

[Ultimately the machine will have to be self- taught and self-teaching. What is the smallest set of heuristics needed to enable a machine to bootstrap itself into 'automated reasoning?]

A 'generalised reasoning machine' would be a marvellous thing, but would it really be accurate to call it an Artificial Mind?

No! There is more to the human mind than the 'working model of the world.' Indeed that is a component that is generated automatically while we sleep. A mind has more than an understanding of how to move about and a vast collection of heuristics.

The missing component is eluded to in the poetic piece "Is There Meaning in Dreams - An Interlude" [RM10].

I describe how the vast bulk of the mind is "mechanical, programmed." (It is the nature of the brain to automate whatever it can - walking, behaviour patterns, conditioned responses.) This is the bit that follows rules - like stepping through a program. This is a 'generalised reasoning machine' though it is one capable of some reprogramming on the fly.

In this vast engine of the mind thought trickles down well-worn paths.

But then I mention something I call 'the magic ball' which can move about in this space - outside this space:

"A small portion of the mind (consciousness) can break its programming - can travel & invent at will."
[Reality Module No.10. p.3]

An AI can follow blindly down the steps of its program, like a mechanical thing - which is what it is.

Machines run programs because they are told to by us the operators. We could tell a computer, for example, to explore a mathematical space - but we would still have to tell it to do this.

Computers have no initiative. They do only what we tell them to do. This is because they are not alive - they are tools which we switch on and switch off.

The current research in AI will give us (maybe in 20 years) a generalised reasoning machine - which will be extremely useful for solving a vast array of problems.

[We could solve some of the problems relating to the limitations of a generalised reasoning machine by having the machine request more data or telling us that "this does not compute." The real problem is when the machine (and the operator) don't realise that there is a knowledge gap. Sure this problem exists with people too (we can be unaware of our ignorance), but it is more serious with computers because we tend to trust too much the results of mechanical calculation.]

But the research will not give us an intelligent machine. That would require a machine with volition - a desire to do its own thing - in essence a conscious curiosity about the data it is working with.

The nature of consciousness remains elusive - and although I am convinced that reasoning (even very high level reasoning) can be automated, I cannot yet see how consciousness can be given to a machine.³

³We could take a conscious entity like a human brain and interface it to a machine.

A machine without consciousness can become artificially intelligent - but a machine would need something akin to consciousness before it could be labelled an artificial mind.

In conclusion - we can automate reasoning, but there isn't a mechanical analogue for consciousness.

[Top of Page]

Related Works

- In This Series -
(1) What Can & Cannot Be Done - The Limits of Futurology (April 2000)
(2) The $20 Computer (April 2000)
(3) Smashing Windows(TM) - The Ascent of Non-linear Thinking (August 2000)
(4) Nu Plastic Yu! (February 2001)
(5) Nu Plastic Yu Tu! (April 2001)
(6) Artificial Minds? (AI Revisited) (August 2001)
(7) Video-On-Demand (June 2002)
(8) Changes (June 2002)
(9) The Implications of Immortality (June 2002)
(10) Cheating in Education (April 2003)

Feedback and Discussions

[Top of Page]

[Return to Home Page]

Copyright © 2001 by Michael F. Green. All rights reserved.

Email.

Last Updated: 8 September 2004