Model of human intelligence/Objective

The goal is to reverse-engineer functional specification of human behavior. It is very similar to goals of cognitive science, but specifically geared towards rapid prototyping of reasonable approximations of human behavior. This page clarifies the objective in detail.

Assumptions

Philosophy of AI is rather complicated. In order to get some actual work done, this project assumes fixed point of view and takes the following assumptions about human intelligence:

• In this project, phrases intelligence, intelligent behavior, and behavior are used interchangeably.
• Intelligence is essentially rationality, an ability to obtain desirable consequences. It is opposite to stupidity which leads to undesirable consequences.
• There are several kinds of rationality. What's desirable, survival or satisfying one's instincts? And for whom, one person, small group, or whole population? This project uses healthy mix of all these kinds of rationality.
• Intelligence is quantitative in nature, it is similar to evolutionary fitness or some kind of success rate. It can be measured or at least estimated to perform at some quantitative level.
• Intelligence as a quantity depends on environment. For purposes of this project, we assume that there is only one possible environment: ecological niche inhabited by humans.
• There are many ways to achieve given level of intelligence. Human intelligence is the one way used by humans. Human intelligence is therefore qualitative in nature. Something either is or isn't human intelligence.
• No simple intelligent algorithm can achieve human-level intelligence, not to mention characteristic human behavior. It is however possible to add up several specialized intelligences to make better performing composite intelligence.
• Human intelligence is assumed to be made up of interconnected components. Main goal of this project is to enumerate and describe these components including their interactions.

Included issues

To avoid creating impression of overly narrow focus of this project, here is a (possibly incomplete) list of issues that are expressly included in scope of this project:

• Obviously human intelligence is within the scope. General-purpose intelligence is included as far as it is used in some component of human intelligence.
• Characteristic human traits even if they do not contribute much to intelligence.
• Description of components of intelligence is major part of this project. This includes interconnections and interactions of components.
• Both brain hardware (neurons) and software (knowledge) is included. Embodied intelligence and cultural knowledge are also addressed. Cultural knowledge is described in somewhat abbreviated form though, just to show how the hardware is used in practice.
• Component description can be either behavioral (black box) or algorithmic (white box).
• Imperfect approximations are also OK, because they can be improved upon by other components.
• Generally answers to all hard questions should be included to make this model implementable without further research.

Excluded issues

In order to control scope of the project, here's a list of issues that are expressly excluded from the scope:

• No super-human and non-human behavior. Only characteristically human behavior is included.
• No need to pass Turing test nor to fool anyone. Unimportant minor behaviors can be left out as far as overall behavior is remarkably human-like.
• No philosophy. See above for concise philosophical assumptions.
• No robots and no software. This is specification. Implementation is for another project.
• No technical design. This is functional specification. Algorithms and data structures are described only if it is easier than describing outward behavior.
• All trivial research (easy questions) that can be conducted with little inventive skill is left for another project. Specifically data sets, tables, and numeric parameters are either attached in separate files/pages, linked to, or entirely excluded from the project.
• No knowledge bases. Knowledge (the software of brain) is described in two ways. First, major ideas/behaviors are explicitly included. Second, large collections of small facts are treated as data sets and usually left out.
• No competing ideas. Only one idea is chosen out of any set of conflicting/redundant ideas. Only one point of view is accepted. Only one consistent specification is developed. One idea approach is used even if it means breaking neutral point of view.
• No need to make compromises to fit in capabilities of current hardware. Compromises to fit in computing capacity of today hardware and/or development of brain-like hardware are topics for another project. No need to design optimizations as far as these can be developed later as part of a specific implementation.

Performance measures

Any specification of intelligence and specifically human intelligence has several performance measures that can be used to evaluate it. These often cannot be precisely measured, but all of them can be roughly estimated. This project aims to maximize all of the benchmarks, but initially completeness is the primary performance goal.

• completeness - How many human behaviors are covered by the specification? This is weighted by importance of these behaviors.
• specificity - What proportion of the specification sticks with human intelligence? This is a measure of redundant content that should be removed from the specification.
• accuracy - How well does specified behavior match actual human behavior?
• implementability - Is the specification sufficiently detailed to allow implementation without further research?
• clarity - How quick and easy is it to comprehend the specification before implementation?
• citation coverage - What proportion of the specification is verified by scientific research?

Open issues

• Embedded links to Wikipedia would explain a lot.