COGNITIVE SYNERGY: A UNIVERSAL PRINCIPLE FOR FEASIBLE GENERAL INTELLIGENCE?

COGNITIVE SYNERGY: A UNIVERSAL PRINCIPLE
FOR FEASIBLE GENERAL INTELLIGENCE?

Ben Goertzel
Novamente LLC
ben@goertzel.org



clear. Without some sort of special
Abstract
assumptions about the environment and goals
Do there exist general principles, which
relevant to an intelligent system, the best way
any system must obey in order to achieve
to achieve general intelligence -- in the sense
advanced general intelligence using feasible
of generally effective goal-achieving behavior,
computational resources? Here we propose
for complex goal/environment combinations --
one candidate: “cognitive synergy,” a
is essentially to carry out some form of brute-
principle which suggests that general
force search of the space of possible behavior-
intelligences must contain different knowledge
control programs, continually re-initiating the
creation
mechanisms
corresponding
to
search as one's actions lead to new
different sorts of memory (declarative,
information. The problem with this approach
procedural,
sensory/episodic,
attentional,
is that it's incredibly computationally
intentional);
and
that
these
different
expensive -- which leads to the question of the
mechanisms must be interconnected in such a
best way to achieve reasonable levels of
way as to aid each other in overcoming
general
intelligence
given
feasible
memory-type-specific
combinatorial
computational resources. Here the picture is
explosions.
less clear mathematically at this point, but,

intuitively, it seems fairly clear that achieving
Keywords: Artificial general intelligence,
useful levels of truly general intelligence
human-level AI, emergence, systems theory
(without special assumptions about the world)

using feasible computational resources is just

not possible.
1. INTRODUCTION

But what if one restricts the scope of

generality, via making appropriate special
What can one say, in general, about
assumptions about the goal/environment
general intelligence? The answer seems to be:
combinations of interest? In this case, one's
a fair bit of interesting mathematics which has
claim for generality of intelligence is less, but
worthwhile philosophical implications, but not
if the restrictions are not too severe, one still
much of direct practical value. The tradition
has a case of interest. But the question is
of Solomonoff induction [1,2], pursued
whether there is anything of elegance and
recently by Juergen Schmidhuber [3], Marcus
scope to say about this case. Plausibly, this
Hutter [4] and others, seems to say what there
case might degenerate into a collection of
is to be said about truly general intelligence.
highly specialized statements about particular
There are many more mathematical details to
classes of goals and environments. On the
be unraveled, but the conceptual picture seems
other hand, maybe there is something sensible
1

---

one can say about "general intelligence in the

Cognitive Synergy Theory, if correct,
everyday world using feasible computational
applies to any AI system demonstrating
resources," as a topic more restrictive than
intelligence in the context of embodied, social
mathematically general intelligence, but less
communication. However, one may also take
restrictive than task-specific intelligence like
the theory as an explicit guide for constructing
chess-playing, car-driving, biological data
AGI systems; and [6] describes one AGI
analysis, etc.
architecture, OpenCogPrime, designed in such

Our goal here is to sketch some ideas
a way.
that we think can serve as the core of a

reasonably general theory of everyday-world
2. ESSENTIAL PRINCIPLES OF
general
intelligence
using
feasible
COGNITIVE SYNERGY THEORY
computational resources. Toward this end we

deal specifically with the case of “multi-
The essential idea of cognitive synergy
memory systems,” which we define as
theory may be expressed in terms of the
intelligent systems whose combination of
following points:
environment, embodiment and motivational

system make it important for them to possess
1. Intelligence, relative to a certain set of
memories that divide into partially but not
environments, may be understood as
wholly distinct components corresponding to
the capability to achieve complex goals
the categories of:
in these environments.

2. With respect to certain classes of goals
• Declarative memory
and environments, an intelligent system
• Procedural memory (memory about
requires a “multi-memory”
how to do certain things)
architecture, meaning the possession of
• Sensory and episodic memory
a number of specialized yet
• Attentional memory (knowledge about
interconnected knowledge types,
what to pay attention to in what
including: declarative, procedural,
contexts
attentional, sensory, episodic and
• Intentional memory (knowledge about
intentional (goal-related). These
the system’s own goals and subgoals)
knowledge types may be viewed as

different sorts of pattern that a system
In [5] we present a detailed argument as to
recognizes in itself and its
how the requirement for a multi-memory
environment.
underpinning for general intelligence emerges
3. Such a system must possess knowledge
from certain underlying assumptions regarding
creation (i.e. pattern recognition /
the measurement of the simplicity of goals and
formation) mechanisms corresponding
environments; but the points made here do not
to each of these memory types. These
rely on that argument. What they do rely on is
mechanisms are also called “cognitive
the assumption that, in the intelligence in
processes.”
question, the different components of memory
4. Each of these cognitive processes, to be
are significantly but not wholly distinct. That
effective, must have the capability to
is, there are significant “family resemblances”
recognize when it lacks the information
between the memories of a single type, yet
to perform effectively on its own; and
there are also thoroughgoing connections
in this case, to dynamically and
between memories of different types.
interactively draw information from
2

---

knowledge creation mechanisms
knowledge, in a way that decreases the
dealing with other types of knowledge
severity of combinatorial explosion.
5. This cross-mechanism interaction must

One prerequisite for cognitive synergy
have the result of enabling the
to work is that each learning mechanism must
knowledge creation mechanisms to
recognize when it is “stuck,” meaning it’s in a
perform much more effectively in
situation where it has inadequate information
combination than they would if
to make a confident judgment about what steps
operated non-interactively. This is
to take next. Then, when it does recognize that
“cognitive synergy.”
it’s stuck, it may request help from other,
6. The activity of the different cognitive
complementary cognitive mechanisms.
processes involved in an intelligent

Next, drilling a little deeper into Point
system may be modeled in terms of the
3 above, one arrives at a number of possible
schematic implication “Context &
knowledge creation mechanisms (cognitive
Procedure ? Goal”, where the Context
processes) corresponding to each of the key
involves sensory, episodic and/or
types of knowledge. Figure 1 below gives a
declarative knowledge; and attentional
high-level overview of the main types of
knowledge is used to regulate how
cognitive process considered in the current
much resource is given to each such
version of Cognitive Synergy Theory,
schematic implication in memory
categorized according to the type of

knowledge with which each process deals.
These points are implicit in the systems theory
Next, Tables 2 and 3 exemplify the memory
of mind given in [7] but are not articulated in
types and cognitive processes from Figure 1 in
this specific form there.
the context of AI systems acting in a simple

Interactions as mentioned in Points 4
virtual world according to the “AGI
and 5 are the real conceptual meat of CST.
Preschool” methodology described in [8]. For
One way to express the key idea here is that
more thorough characterizations of these ideas,
most AI algorithms suffer from combinatorial
see [7].
explosions: the number of possible elements to

be combined in a synthesis or analysis is just

too great, and the algorithms are unable to
3. THE COGNITIVE SCHEMATIC
filter through all the possibilities, given the

lack of intrinsic constraint that comes along
Point 6 in the above summary of Cognitive
with a “general intelligence” context (as
Synergy Theory describes how the various
opposed to a narrow-AI problem like chess-
cognitive processes involved in intelligence
playing, where the context is constrained and
may be understood to work together via the
hence restricts the scope of possible
“cognitive schematic”
combinations that needs to be considered). In

an AGI architecture based on CST, the
Context & Procedure ? Goal <p>
different learning mechanisms must be

designed specifically to interact in such a way
This formula may interpreted to mean “If
as to palliate each others’ combinatorial
the context C appears to hold currently, then if
explosions – so that, for instance, each
I enact the procedure P, I can expect to achieve
learning mechanism dealing with a certain sort
the goal G with certainty p.” The system is
of knowledge, must synergize with learning
initially supplied with a set of high-level goals
mechanisms dealing with the other sorts of
such as “get rewarded by my teacher”, “learn
new things” and so forth; and it then uses
3

---

inference
(guided
by
other
cognitive
and subgoals. In the following will also use the
mechanisms) to refine these initial goals into
shorthand
more specialized subgoals. As noted above,

we use the term “intentional knowledge” to
C & P ? G <p>
refer to the system’s knowledge of its goals





Figure 1. High-level overview of the key cognitive dynamics considered in the current version of Cognitive
Synergy Theory. Cognitive Synergy Theory in its current form describes the behavior of a system as it pursues
a set of goals, which are then refined by inference, aided by other processes. Terms like “inference” are used
very broadly here; for instance there is no commitment to explicit use of a logic engine and, from the point of
view of a high-level description like this diagram, inference could just as well be carried out as an emergent
process resulting from the dynamics of an neural net system. At each time the system chooses a set of procedures
to execute, based on its judgments regarding which procedures will best help it achieve its goals in the current
context. These procedures may involve external actions (e.g. involving conversation, or controlling an agent in
a simulated world) and/or internal cognitive actions. In order to make these judgments it must effectively
manage declarative, procedural, episodic, sensory and attentional memory, each of which is associated with
specific algorithms and structures as depicted in the diagram. There are also global processes spanning all the
forms of memory, including the allocation of attention to different memory items and cognitive processes, and
the identification and reification of system-wide activity patterns (the latter referred to as “map formation”).


4

---

Knowledge Type
Virtual Agent Example(s)
Declarative
• The red ball on the table is larger than the blue ball
on the floor
• Bob becomes angry quickly
• Ball roll. Blocks don’t.
• Jim knows Bob is not my friend.
Procedural
• A procedure for retrieving an item from a distant
location
• A procedure for spinning around in a circle
• A procedure for stacking a block on top of another
one
• A procedure for repeatedly asking a question in
different ways until an acceptable answer is obtained
Sensory
• The appearance of Bob’s face
• The specific array of objects on the floor under the
table
Episodic
• The series of actions Bill did when he built a tower
on the floor yesterday
• The episode in which Bill and Bob repeatedly threw
a ball back and forth between each other
• The series of actions I just took, between getting up
from the chair and Bob saying “good”
Attentional
• The set of objects that seem to be important in the
context of the game Bob and Bill are playing
• The set of words and phrases that are associated with
Bob being happy with me while we walk around
together
Intentional
• The goal of making Bob say positive things
• The goal of making a tower that does not fall down
easily
• The goal of getting Jim to answer my question

Table 1. Examples of the key knowledge types in the context of virtual agent control

5

---

Cognitive Process
Virtual Agent Example
Inference
•
Tall thin blocks, when stood upright, are less likely to topple
over if placed next to each other
•
Bob hates cursing, and Jim is Bob’s friend, and friends often
have similar likes and dislikes, so Jim probably hates cursing
Procedure Learning
•
Learning a procedure for crawling on the floor, based on
imitation of what others do when they describe themselves as
“crawling”, plus reinforcement from others when they find
one’s imitation accurate
•
Learning a procedure embodying some combination of
functional and visual features that predicts whether some
entity is considered a toy or not
Attention allocation
•
Pictures of women are associated with Bob’s happiness, and
Bob’s happiness is associated with getting reward, therefore
pictures of women are associated with getting reward
•
Asking for help is surprisingly often a precursor to getting
reward when Jane is around; so when a reward is gotten when
Jane is around, a little extra attention should be given to
ongoing improvement of the processes that help in the
mechanics of asking for help
Goal refinement
•
The goal of making Jim happy, seems to often be achieved by
the goal of creating sculptures Jim likes, and Jim likes
complicated sculptures; thus I adopt the goal of creating
complicated sculptures when Jim is around
Declarative pattern mining
•
Tall thin blocks, when stood upright, are likely to topple over
Sensory pattern recognition
•
When Jim builds a castle out of blocks, he identifies some
portions of the castle as “towers” and others as “walls”; it’s
necessary to visually identify which portions of each castle
correspond to these descriptors
•
It’s also necessary to visually identify the castle as a whole
versus the table, floor or other base it’s resting on
Simulation
•
Using an internal simulation world to experiment with
building various towers rapidly, at a pace faster than is
possible in the online simulation world where humans
participate
•
Using an internal simulation world containing a simulation of
Bob and Jim, to simulate what Bob will know about what
you’re doing if you hide behind Jim and build a tower of
blocks
Concept creation
•
The concept of an unstable structure
•
The concept of an irritable person
•
The concept of a happy occasion
Map formation
•
The set of all knowledge items associated with Bob being in a
good mood (which may then be used to form a new concept)
•
The set of all knowledge items associated with (running,
walking or crawling) races

Table 2. Examples of the key cognitive processes in the context of virtual agent control
6

---

In general, the cognitive schematic leads
G, given fixed P and C. Simulation may
to a conceptualization of the internal action of
also be used for this purpose.
an intelligent system as involving two “key
• Goal refinement is used to create new
learning processes”:
subgoals G to sit on the right hand side

of instances of the cognitive schematic
1. Estimating the probability p of a posited
• Concept formation and map formation
C & P ? G relationship
are useful for choosing G and for fueling
2. Filling in one or two of the variables in
goal refinement, but especially for
the cognitive schematic, given
choosing C (via providing new
assumptions regarding the remaining
candidates for C). They can also be
variables, and directed by the goal of
useful for choosing P, via a process
maximizing the probability of the
called “predicate schematization” that
cognitive schematic
turns logical predicates (declarative

knowledge) into procedures.
... or, to put it less technically:



On the other hand, where analysis is
1. Evaluating conjectured relationships
concerned:
between procedures, contexts and goals

(“analysis”)
• Inference, acting on declarative
2. Conceiving novel possible relationships
knowledge, can be useful for estimating
between procedures, contexts and goals
the probability of the implication in the
(“synthesis”); and when necessary
schematic equation, given fixed C, P and
creating new procedures and contexts,
G. Episodic knowledge can also be
via leveraging prior knowledge or as a
useful in this regard, via enabling
last resort via trial and error
estimation of the probability via simple
experimentation
similarity matching against past

experience. Simulation may also be
Given this conceptualization, we can see
used: multiple simulations may be run,
that, where synthesis is concerned,
and statistics may be captured therefrom.

• Procedural knowledge, mapped into
• Procedural knowledge, and procedural
declarative knowledge and then acted on
learning methods can be useful for
by inference, can be useful for
choosing P, given fixed C and G.
estimating the probability of the
Simulation may also be useful, via
implication C & P ? G, in cases where
creating a simulation embodying C and
the probability of C & P1 ? G is
seeing which P lead to the simulated
known for some P1 related to P
achievement of G
• Inference, acting on declarative or
• Declarative knowledge, and associated
sensory knowledge, can be useful for
knowledge creation methods, can be
estimating the probability of the
useful for choosing C, given fixed P and
implication C & P ? G, in cases where
G (also incorporating sensory and
the probability of C1 & P ? G is
episodic knowledge as useful).
known for some C1 related to C; and
Simulation may also be used for this
similarly for estimating the probability
purpose.
of the implication C & P ? G, in cases
• Inference, acting on declarative
where the probability of C & P ? G1 is
knowledge, can be useful for choosing
known for some G1 related to G
7

---

8

---

Table 3. Synergies between the cognitive processes shown in Figure 1.
9

---

Map formation and concept creation can be
significant
inventiveness
and
may
be
useful indirectly in calculating these probability
approached in many different ways. The
estimates, via providing new concepts that can
primary approach we have pursued involves the
be used to make useful inference trails more
OpenCogPrime
software
design,
which
compact and hence easier to construct.
introduces specific algorithms for each of the
The key role of attentional knowledge in the
capabilities mentioned in Figure 1, together with
overall functioning of intelligent systems as
specific mechanisms for realizing the synergies
described by CST must be emphasized. In any
in Table 1. The specifics of how
real-world context, a system will be presented
OpenCogPrime manifests these synergies are
with a huge number of possibly relevant
discussed further in [7].
analysis and synthesis problems. Choosing

which ones to explore is a difficult cognitive
References
problem in itself – a problem that also takes the

form of the cognitive schematic, but where the
[1] Solomonoff, Ray. 1964. "A Formal Theory
procedures are internal rather than external.
of Inductive Inference, Part I"
Thus this problem may be addressed via the
Information and Control, Vol 7, No. 1 pp 1-22,
analysis and synthesis methods describe above.
March 1964.
This is the role of attentional knowledge.
[2] Solomonoff, Ray. 1964. "A Formal Theory

Finally, one way to see the essential role
of Inductive Inference, Part II"
of synergy in intelligence as modeled by CST, is
Information and Control, Vol 7, No. 2 pp 224-
to observe that sometimes the best way to
254, June 1964
handle the schematic equation will be to fix only
[3] Schmidhuber, J. (2006). Gödel machines:
one of the terms. For instance, if we fix G,
Fully Self-Referential Optimal
sometimes the best approach will be to
Universal Self-Improvers. In B. Goertzel and C.
collectively learn C and P. This requires either
Pennachin, eds.: Artificial
a procedure learning method that works
General Intelligence, p. 119-226, 2006.
interactively with a declarative-knowledge-
[4] Hutter, Marcus (2005). Universal Artificial
focused concept learning or reasoning method;
Intelligence: Sequential Decisions
or a declarative learning method that works
based on Algorithmic Probability. Springer,
interactively with a procedure learning method.
2005

[5] Goertzel, Ben (2009). The Embodied

Communication Prior. Dynamical Psychology
4. ENUMERATION OF CRITICAL
[6] Goertzel, Ben (2009). OpenCog Prime: A
SYNERGIES
Cognitive Synergy Based Architecture for

General Intelligence. Submitted for publication

Referring back to Figure 1, and
[7] Goertzel, Ben (2006). The Hidden Pattern.
summarizing many of the ideas in the previous
Brown Walker
section, Table 1 enumerates a number of
[8] Goertzel, Ben and Stephan Vladimir Bugaj
specific ways in which the cognitive processes
(2009). AGI Preschool. Proceedings of AGI-
mentioned in the Figure may synergize with one
09, Atlantis Press.
another, potentially achieving dramatically

greater efficiency than would be possible on

their own.

Of course, realizing these synergies on the
practical algorithmic level will require
10

---