When we talk about artificial intelligence, human intelligence, or any other form of intelligence, what do we mean by the term intelligence in a general sense? What would you call intelligent and what not? In other words, how do we define the term intelligence in the most general possible way?
I'm going to preface this answer by noting that persons much smarter than myself have treated this subject in some detail. That said, as far as I can discern:
When we talk about intelligence we're referring to problem solving strength in relation to a problem, relative to the strength of other intelligences.
This is a somewhat game-theoretic conception, related to rationality and the concept of the rational agent. Regarding intelligence in this manner may be unavoidable. Specifically, we could define intelligence as the ability to understand a problem or solution or abstract concepts, but we can't validate that understanding without testing it. (For instance, I might believe I grasp a mathematical technique, but the only way to determine if that belief is real or illusory is to utilize that technique and evaluate the results.)
The reason games like Chess and Go have been used as milestones, aside from longstanding human interest in the games, is that they provide models with simple, fully definable parameters, and, in the case of Go at least, have complexity akin to nature, by which I mean unsolvable/intractable. (Compare to strength at Tic-Tac-Toe, which is trivially solved.)
However, we should consider a point made in this concise answer to a question involving the Turing Test:
"...is [intelligence] defined purely by behaviour in an environment, or by the mechanisms that arrive at that behaviour?"
This is important because Google just gave control over data center cooling to an AI. Here it is clearly the mechanism itself that demonstrates utility, but if we call that mechanism intelligent, for intelligence to have meaning, we still have to contend with "intelligent how?" (In what way is it intelligent?) If we want to know "how intelligent?" (its degree of utility) we still have to evaluate its performance in relation to the performance of other mechanisms.
(In the case of the automata controlling the air conditioning at Google, we can say that it is more intelligent than the prior control system, and by how much.)
Because we're starting to talk about more "generalized intelligence", defined here as mechanisms that can be applied to a set of problems, (I include minimax as a form of "axiomatic intelligence" and machine learning as a form "adaptive intelligence"), it may be worthwhile to expand and clarify the definition:
Intelligence is the problem solving strength of a mechanism in relation to a problem or a set of problems, relative to the strength of other mechanisms.
or, if we wanted to be pithy:
Intelligence is as intelligence does (and how well.)
Over the years, many people have attempted to define intelligence, so there are many definitions of intelligence, but most of them are not formalized. For a big collection of definitions, see the paper A Collection of Definitions of Intelligence (2007) by Shane Legg and Marcus Hutter.
In an attempt to formally define intelligence, so that it comprises all forms of intelligence, in the paper Universal Intelligence: A Definition of Machine Intelligence (2007), the same Legg and Hutter, after having researched the previously given definitions of intelligence, define intelligence as follows
Intelligence measures an agent's ability to achieve goals in a wide range of environments
This definition apparently favors systems that are able to solve more tasks (i.e. AGIs) than systems that are only able to solve a specific task (i.e. narrow AIs), but, according to Legg and Hutter, it should summarise the main points of the previously given definitions of intelligence, so it should be a reasonable and quite general definition of intelligence. Moreover, properties associated with intelligence, like the ability to learn, should be emergent, i.e. in order to achieve goals in a wide range of environments you also need the ability to learn.
In my blog post On the definition of intelligence, I also talk about this definition, but I suggest that you read the mentioned papers if you are interested in all details. This video by Marcus Hutter could also be useful and interesting.
This is an important question for AI – maybe the most important of all – for the research field of Artificial Intelligence. I mean if AI is science, then its experiments will be empirically testable. There has to be a way to decide pass or fail. So what are the tests for intelligence? Before you even design a test, you need a clear idea of what intelligence amounts to, otherwise how could you design a competent test for it?
Sure, I'm part of the research and development project known as Building Watertight Submarines, and sure, I'm totally confident my submarine is watertight, but I have no idea how to test whether it is or not because I don't know what "watertight" means. This whole idea is absurd. But ask AI what "intelligence" means. The answers you get, on analysis, are almost the same as the submarine example.
Base Answer - Behavior
The word (idea, concept) "Intelligence" is usually defined by AI in terms of behavior. I.e. the Turing test approach. A machine is intelligent if it behaves in a way that, were a human to behave in that same way, the human would be said to be performing an action that required human intelligence.
Problem 1: player pianos are intelligent. Playing a Scott Joplin tune obviously requires intelligence in a human.
Problem 2. If a machine passes the test, it only shows that the machine is "intelligent" for the tested behaviors. What about untested behaviors? This is actually a life-and-death problem today with self-driving vehicle AI control systems. The AI systems are acceptably good at driving a car (which obviously requires human intelligence) in specific environments, e.g. freeways with well-marked lanes, no tight corners, and a median barrier separating the two directions. But the systems go disastrously wrong in "edge cases" – unusual situations.
Problem 3. Who would put their child on a school bus driven by a robot that had passed the Turing test for driving school buses? What about a storm when a live power line falls across the road? Or a twister in the distance is coming this way? What about a thousand other untested possibilities? A responsible parent would want to know (a) what are the principles of the internal processes and structures of human intelligence, and (b) that the digital bus driver had adequately similar internal processes and structures – i.e., not behavior but the right inner elements, the right inner causation.
Desired answer – inner principles
I would want to know that the machine was running the right inner processes and that it was running these processes (algorithms) on the right inner (memory) structures. Problem is, no one seems to know what the right inner processes and structures of human intelligence are. (A huge problem to be sure – but one that hasn't held AI back – or self-driving system developers - one bit.) The implication of this is that what AI ought to be doing now is working out what are the inner processes and structures of human intelligence. But it's not doing this – rather, it's commercializing its flawed technology.
Elements of a definition – 1. Generalization
We do know some things about human intelligence. Some tests really do test whether a machine has certain properties of the human mind. One of these properties is generalization. In his 1950 paper, Turing, as a sort of joke, gave a really good example of conversational generalization: (The witness is the machine.)
Interrogator: In the first line of your sonnet which reads 'Shall I compare thee to a summer's day', would not 'a spring day' do as well or better?
Witness: It wouldn't scan.
Interrogator: How about 'a winter's day' That would scan all right.
Witness: Yes, but nobody wants to be compared to a winter's day.
Interrogator: Would you say Mr. Pickwick reminded you of Christmas?
Witness: In a way.
Interrogator: Yet Christmas is a winter's day, and I do not think Mr. Pickwick would mind the comparison.
Witness: I don't think you're serious. By a winter's flay one means a typical winter's day, rather than a special one like Christmas.
Current AI has nothing that comes even remotely near being able to generalize like this. Failure to generalize is regarded as perhaps the greatest failing of current AI. The ability to generalize would be one part of an adequate definition of "intelligence". But what generalization amounts to would need to be explicated.
The problem of generalization, also, is behind several the severe philosophical objections to AI theory, including the frame problem, the problem of common-sense knowledge, and the problem of combinatorial explosion.
Elements of a definition – 2. Perception
Sensory perception is fairly obviously fundamental to human learning and intelligence. Data (in some form) is emitted by the human senses then processed by the central system. In the computer, binary values exit the digital sensor and travel to the machine. However, nothing in the values themselves indicates what was sensed. Yet the only thing the computer gets is the binary values. How could the machine ever come to know what is sensed? (The classic Chinese room argument problem.)
So another element of human-like intelligence is the ability to perceive in a human-like way. What "human-like way" means here is that the machine processes sensory input using the same principles that apply in human perception. The problem is that no one seems to know how a semantics (knowledge) can be built from the data emitted by digital sensors (or organic senses). But still, human-like perception needs to be an element of an adequate definition of "intelligence".
Once AI gets these two issues sorted out – generalization and perception – then it will probably, hopefully, be well on the way to realizing its original goal of almost 70 years past – building a machine with (or that could acquire) a human-like general intelligence. And maybe the principles of generalization and the principles of perception are one and the same. And maybe there is actually only one principle. It shouldn't be assumed that the answers are complex. Sometimes the hardest things to understand are the most simple.
So the question "What do we mean when we say "intelligence"? is really important to AI. And the conclusion is that AI ought to replace its current behavioral definition of "intelligence" with one that includes the human elements of generalization and perception. And then get on and try to work out the operating principles, or principle, of both of these.
Intelligence is the ability to weave together various concepts and associations into a meaningful whole; filtering, adding and rejecting appropriately various ideas from personal knowledge and experience. Then effectively reflecting these ideas back to a questioner to affirm understanding and comprehension, allowing a conversation to proceed effectively towards a mutually beneficial conclusion.
The most general definition of the term intelligence that is both terse and exact is this.
The collection of behavioral features resident in some entity where the entity sustainably succeeds in specific pursuits while avoiding specific losses in a particular range of environmental conditions.
These are examples of failures in exhibiting intelligence per the above definition, demonstrating the importance of each phrase.
- Some behavioral features that are intelligent but an overall system behavior that is not. For example, a rocket that reaches an altitude but cannot enter orbit or a turtle that can retract its head but cannot catch a bug.
- The intelligence is scattered across entities such that each individual entity does not exhibit intelligence, such as a bee or a single.
- Intelligence is exhibited momentarily but disintegrates over time, not adapting to changing conditions, or is not sufficiently reliable to fill a practical role.
- The entity can attain a goal but such success is nullified by loss accumulated during their attainment.
- The entity can avert loss, but cannot reliably succeed in the pursuit of its objectives.
- An entity that can adapt to any environmental condition and act intelligently in every one does not exist. Human intelligence is limited to specific scenarios and shock and confusion result when overburdened and super-intelligence is, as of this writing, conjecture without empirical evidence or theoretical proof.
Notice four things in this definition.
- Optimality is not required. Only better than random behavior is required.
- Although the entity tested for intelligence may interface with its environment only through data sets and test metrics, these are its environment.
- Time is necessarily involved. In a simple case, an artificial network exhibits intelligence only in its ability to exhibit behavior that was previously learned to be sufficient. Such can only retain intelligence by adjusting its training or in an environment where adaptation to new patterns is not required.
- Cognition is not required but cognition certainly augments the range of objectives that can be reliably pursued and the ability of the entity to detect danger and more proactively avert loss.