Status: Draft.
Version: 1.0
Author: Leo Meyer, leo@leomeyer.de
Note: The thoughts expressed here seem so obvious that they have probably been mentioned elsewhere. However, they serve as theoretical background for the Associator project and are therefore necessary to clarify my own views. My ignorance of related literature additionally relieves me of cumbersome compilations of bibliographies, as well as of blaming others for any faults in my thoughts. If you find any of these, assume that there is a flaw in my reasoning and kindly notify me of it.
This article discusses the following topics:
The primary cause why we want help from a "thinking machine" is that we are dissatisfied with our own intelligence. (At least this is true for me.) I am ignorant of many interesting facts and abilities, and my work continually demands that I should learn - and more important, remember - those to accomplish the things I want to do. But unfortunately vita brevis, ars longa. As we have an ardent desire for truth, and do not like to commit mistakes, we find the notion extremely discomforting that our thinking, reasoning and acting constantly depends on incomplete knowledge resources and is conducted using imperfect means. But there is no time to read and memorize all newspapers to decide whom we are going to elect next, for example. Nor is there time to become perfect in everything. We humans have to accept that we are limited both in time and memory. The boundary of our knowledge is the last frontier separating us from God.
A secondary cause is that we want to delegate "mechanical" tasks, which nevertheless involve some kind of feedback and decision-making, to cheap servants, which will in turn give us more time to do what we really want to do, namely, enjoy the creative faculty of our minds.
We assume that artificial intelligence, as created by humans, will be greatly superior to our individual natural intelligence. But probably, if there was a machine behaving exactly like an ordinary human and occasionally exhibiting intelligent behaviour, it would be classified by most of us as rather dumb. If we then undertake the task of making an intelligent machine, it should at least be an Albert Einstein in physics, a Mozart on the pianoforte, and a Bill Gates in marketing. To express this idea in more formal terms:
Assumption 1: We want to build a mechanical or electronic system which can emulate the human ability to acquire, store, use and reproduce knowledge. If possible, this system should not be as susceptible to errors as humans are. Specifically, it should not forget knowledge once acquired, and it should use all knowledge correctly under all circumstances.
But what exactly do we mean by the term "intelligence", and what makes it "artificial" - obviously in contrast to something we could call "natural intelligence"?
I suggest that we should not define intelligence as something specifically human. Otherwise we would not be able to create AI in the first instance, because all intelligent agents would necessarily be human, or would simply not be intelligent as per definition. Furthermore, our intention goes towards making an "intelligent machine", not a being; because we live under the Frankensteinian illusion that machines we created can also be controlled by us. We will see if this holds true for an intelligent machine as well.
This raises the dilemma that we must assume that human intelligent behaviour consists of operations which can be performed by a machine, and consequently, from this point of view, the human brain (or at least its faculty for intelligent behaviour) would be a machine. If we humbly admit this (and I doubt not there will be much discussion about this point), we must only decide on a mathematical model which describes this faculty of the human brain so that we can implement it into a computer system, and there is our artificial intelligence.
Many discussions on AI rely on logic as the principal means of creating artificial intelligence. This is a misconjecture resulting from overestimating the power of logic and from not properly understanding the way the human brain functions.
Quickly summing up the assumptions of the logicians:
1. A statement can be either true or false.
2. The universe consists of separate entities and their relations.
3. A state of the universe (for example, the current state) can be accurately and completely described using statements about entities and their relations.
Theorem 1.1: A system of coherent logical statements becomes contradictory when it reaches sufficient complexity.
Proof: Suppose there is a set of n statements which are known to be either true or false. The nature of these statements is such that they allow reasoning, that is the formation of new statements according to derivation rules and the assignment of definite truth values to the derived statements.
Whenever a statement is added, the logical system, to remain free of contradictions, must check
1. whether the new statement contradicts any of the previous statements,
2. whether the application of all derivation rules to all combinations of the new and previous statements leads to a contradiction with the previous statements,
3. whether the resulting statements of the derivation process, when subjected to the same procedure as described in step 2, lead to a contradiction, and so on.
It can be easily seen that the effort or cost of integrity checks grows beyond calculation limits as n grows larger. The system therefore eventually reaches a point where it cannot check all combinations of statements anymore. At this point it becomes error-prone, and an error (or a contradiction) in a reasoning system renders the whole system inoperable.
q.e.d.
Personally I have no doubt that the number of statements required to produce anything similar to intelligent behaviour by far exceeds this magical limitation point.
The second argument against basing AI entirely on logic is that the human brain is not relying on logic for its intelligent behaviour as well. Logic in humans is something which occurs only after a language has been learnt, and its proper acquisition so difficult that many persons never accomplish it in their whole life; anyhow, they act intelligently, therefore intelligence must be more than mere logical reasoning.
Presumption 1.1: Intelligence and language are closely related.
Presumption 1.2: Intelligence and memory are closely related.
Presumption 1.3: Intelligence emerges as language is processed within a memory.
First, then, we have to define what we mean by knowledge and how it relates to data.
Def 1.1: Data. Data is anything which can be stored by a computer.
Def 1.2: Information. Information is data which has been classified by an intelligent agent as such.
Def 1.3: Knowledge. Knowledge is information retained by intelligent agents in the form of thoughts.
Def 1.4: Thought. A thought is an idea which can be expressed in a language.