Levels of Organization in General Intelligence is a draft of a paper by Eliezer Yudkowsky, to appear in Ben Goertzel and Cassio Pennachin, (eds.) "Artificial General Intelligence". The draft may differ from the final paper.
The model of intelligence presented
in this chapter - "Deliberative General Intelligence" or "DGI" - requires
five distinct layers of organization, each layer built on top of the underlying
layer.
-
The bottom layer is source code and data structures - complexity
that is manipulated directly by the programmer. The equivalent layer
for humans is neurons and neural circuitry.
-
The next layer is sensory modalities. In humans, the archetypal
examples of sensory modalities are sight, sound, touch, taste, smell, and
so on; implemented by the visual areas, auditory areas,
et cetera. In biological brains, sensory modalities come the closest
to being "hardwired"; they generally involve clearly defined stages of
information-processing and feature-extraction, sometimes with individual
neurons playing clearly defined roles. Thus, sensory modalities are
some of the best candidates for processes that can be directly coded by
programmers without rendering the system crystalline and fragile.
-
The next layer is concepts. Concepts (also sometimes known
as "categories", or "symbols") are abstracted from our experiences.
Abstraction reifies a perceived similarity within a group of experiences.
Once reified, the common quality can then be used to determine whether
new mental imagery satisfies the quality, and the quality can be
imposed
on a mental image, altering it. Having abstracted the concept "red",
we can take a mental image of a non-red object (for example, grass) and
imagine "red grass". Concepts are patterns that mesh with sensory
imagery; concepts are complex, flexible, reusable patterns that
have been reified and placed in long-term storage.
-
The next layer is thoughts, built from structures of concepts.
By imposing concepts in targeted series, it becomes possible to build up
complex mental images within the workspace provided by one or more sensory
modalities. The archetypal example of a thought is a human "sentence"
- an arrangement of concepts, invoked by their symbolic tags, with internal
structure and targeting information that can be reconstructed from a linear
series of words using the constraints of syntax, constructing a complex
mental image that can be used in reasoning. Thoughts (and their corresponding
mental imagery) are the disposable one-time structures, built from reusable
concepts, that implement a non-recurrent mind in a non-recurrent world.
-
Finally, it is sequences of thoughts that implement deliberation
- explanation, prediction, planning, design, discovery, and the other activities
used to solve knowledge problems in the pursuit of real-world goals.
Although the five-layer model is central to the DGI theory of intelligence,
the rule of Necessary But Not Sufficient still holds. An AI
project will not succeed by virtue of "implementing a five-layer model
of intelligence, just like the human brain". It must be the right
five layers. It must be the right modalities, used in the
right
concepts, coming together to create the right thoughts seeking out
the right goals. (An AI might use different modalities, but
will still need a right set of modalities.)
The five-layer model of deliberation is not inclusive of everything
in the DGI theory of mind, but it covers substantial territory, and can
be extended beyond the deliberation superprocess to provide a loose sense
of which level of organization any cognitive process lies upon. Observing
that the human body is composed of molecules, proteins, cells, tissues,
and organs is not a complete design for a human body, but it is nonetheless
important to know whether something is an organ or a protein. Blood,
for example, is not a prototypical tissue, but it is composed of cells,
and is generally said to occupy the tissue level of organization of the
human body. Similarly, the hippocampus, in its role as a memory-formation
subsystem, is not a sensory modality, but it can be said to occupy the
"modality level": It is brainware (a discrete, modular chunk of neural
circuitry); it lies above the neuron/code level; it has a characteristic
tiling/wiring pattern as the result of genetic complexity; it interacts
as an equal with the subsystems comprising sensory modalities.
Generalized definitions of the five levels of organization might be
as follows:
-
Code-level, hardware-level: No generalized definition is needed,
except that the biological equivalent is the neural level or wetware
level.
-
Modality-level: Subsystems which, in humans, derive their
adaptive complexity from genetic specification - or rather from the genetic
specification of an initial tiling pattern and a self-wiring algorithm,
and from exposure to invariant environmental complexity. The AI equivalent
is complexity which is known in advance to the programmer and which is
directly specified through programmer efforts. Full systems on this
level are modular parts of the cognitive supersystem - one of a large but
limited number of major parts making up the mind. Where the system
in question is a sensory modality or a system which clearly interrelates
to the sensory modalities and performs modality-related tasks, the system
can be referred to as modality-level. Similarly, a subsystem
or subprocess of a major modality-level system, or a minor function
of such a subsystem, may also be referred to as modality-level. Where
this term is inappropriate, because a subsystem has little or no relation
to sensory modalities, the subsystem may be referred to as brainware.
-
Concept-level: Concepts are cognitive objects which are placed
in long-term storage, and reused as the building blocks of thoughts.
The generalization for this level of organization is learned complexity:
cognitive content which is derived from the environment and placed in long-term
storage, and which thereby becomes part of the permanent reservoir of complexity
with which the AI challenges future problems. The term
concept-level
might optionally be applied to any learned complexity that resembles categories;
i.e., learned complexity that interacts with sensory modalities and acts
on sensory modalities. Regardless of whether they are conceptlike
(an issue considered later), other examples of learned complexity include
declarative beliefs and episodic memories.
-
Thought-level: A thought is a specific structure of combinatorial
symbols which builds or alters mental imagery. The generalizable
property of thoughts is their immediacy. Thoughts are not
evolved/programmed brainware, or a long-term reservoir of learned complexity;
thoughts are constructed on a moment-by-moment basis. Thoughts make
up the life history of a non-recurrent mind in a non-recurrent universe.
The generalized thought level extends beyond the mentally spoken sentences
in our stream of consciousness; it includes all the major cognitive events
occurring within the world of active mental imagery, especially events
that involve structuring the combinatorial complexity of the concept level.
-
Deliberation, like the code level, needs no generalization.
Deliberation describes the activities carried out by patterns of thoughts.
The patterns in deliberation are not just epiphenomenal properties of thought
sequences; the deliberation level is a complete layer of organization,
with complexity specific to that layer. In a deliberative AI, it
is patterns of thoughts that plan and design, transforming abstract high-level
goal patterns into specific low-level goal patterns; it is patterns of
thoughts that reason from current knowledge to predictions about unknown
variables or future sensory data; it is patterns of thoughts that reason
about unexplained observations to invent hypotheses about possible causes.
In general, deliberation uses organized sequences of thoughts to solve
knowledge problems in the pursuit of real-world goals.
Even for the generalized levels of organization, not everything fits cleanly
into one level or another. While the hardwired-learned-invented trichotomy
usually matches the modality-concept-thought trichotomy, the two are conceptually
distinct, and sometimes the correspondence is broken. But the levels
of organization are almost always useful - even exceptions to the rule
are more easily seen as partial departures than as complete special cases.
- 1: Other human modalities include, e.g., proprioception and vestibular
coordination.
- 2: Environmental
complexity of this type is reliably present and is thus "known in advance"
to the genetic specification, and in some sense can be said to be a constant
and reliable part of the genetic design.
- 3: The
term "brainware" is not necessarily anthropomorphic, since the term "brain"
can be extended to refer to nonbiological minds. The biology-only
equivalent is often half-jokingly referred to as wetware, but the
term "wetware" should denote the human equivalent of the code level, since
only neurons and synapses are actually wet.