THE EMBODIED COMMUNICATION PRIOR: A CHARACTERIZATION OF GENERAL INTELLIGENCE IN THE CONTEXT OF EMBODIED SOCIAL INTERACTION

THE EMBODIED COMMUNICATION PRIOR: A
CHARACTERIZATION OF GENERAL INTELLIGENCE IN
THE CONTEXT OF EMBODIED SOCIAL INTERACTION

Ben Goertzel
Novamente LLC
ben@goertzel.org



Abstract
1. INTRODUCTION


We outline a general conceptual definition
“Intelligence,”
like
most
folk-
of real-world general intelligence that avoids
psychology concepts, is slippery with multiple
the twin pitfalls of excessive mathematical
overlapping
meanings,
and
defies
generality, and excessive anthropomorphism..
straightforward formalization. Legg and Hutter
Drawing on prior literature, a definition of
[1] have reviewed over 70 published definitions
general intelligence is given, which defines the
of intelligence, and have proposed a fully
latter by reference to an assumed measure of the
formalized definition of their own [2]. One
simplicity of goals and environments. The novel
thing that becomes clear from their survey is
contribution presented is to gauge the simplicity
that different definitions of intelligence have
of an entity in terms of the ease of
often been formulated with different goals in
communicating it within a community of
mind. Here our intent is to say something new
embodied agents (the so-called Embodied
about the topic by formulating the goal in a
Communication Prior or ECP). Augmented by
novel way. We aim specifically to say
some further assumptions about the statistical
something about general intelligence in the
structure of communicated knowledge, this
sense of “embodied, intercommunicating beings
choice is seen to lead to a model of intelligence
who display general intelligence in the context
in terms of distinct but interacting memory and
of their interactions with each other and the
cognitive subsystems dealing with procedural,
world.” This is a narrower scope than many of
declarative, sensory/episodic, attentional and
the mathematical or philosophical definitions
intentional knowledge. A sister paper then
that have been proposed, but a broader scope
extends these ideas to yield a “Cognitive
than approaches that are tied specifically to
Synergy Theory” that suggests specific
intelligence that is closely humanlike.
conclusions for the architecture of artificial
We will begin with a very general, semi-
general intelligences, based on the ECP.
formalized definition of intelligence that is

largely in the spirit of Legg and Hutter [2], and

also in the spirit of the earlier semi-formalized
Keywords: Artificial general intelligence,
definition from [3]. The intuition underlying
human-level AI, Universal AI, learning theory,
this definition is that “intelligence is the ability
embodiment
to achieve complex goals in complex

environments.” To capture this more precisely,

we may assess intelligence by averaging
1

---

achievability (by the system) over the space of
determine its next action, it will be maximally
all (goal, environmental situation) pairs, where
likely to achieve its goals. After taking a single
the achievability of the pair is weighted by a
action, it then re-evaluates and chooses a new
coefficient proportional to the simplicity of the
program P to execute; etc. This is a great
pair according to the ECP. This begs the
approach if one has the computational resources
question of how to measure simplicity – but we
to perform a thorough search of program space
will defer this to the following section of the
at each time step. Unfortunately AIXI requires
paper, as it’s one of the key original themes
infinitely much computational resources, but
presented here.
there is a related approach called AIXItl that
So: given a system M, a goal G and a
restricts its search to programs with specifically
situation S, let ach(M,G,S) denote the quality
bounded length and runtime, whose resource
with which M can achieve G in S; and let
requirements are merely infeasibly humongous
simp(G,S) denote the simplicity of the pair
rather than infinite. This kind of approach has
(G,S). Then, we may conceive the intelligence
very high general intelligence, but very poor
of M as the sum over (G,S) pairs of
efficient intelligence. Yet, our strong suspicion

is that if one measures simplicity using standard
ach(M,G,S) simp(G,S)
computational models, there’s probably not any

way to do significantly better.
Fully formalizing this along the lines of [2]

The trick to real-world general
would be possible but would be out of place
intelligence, we suggest, lies in the simplicity
here as this is a qualitative paper that won't be
measure. If one assumes a simplicity measure
pursuing rigorous proofs.
that is biased to certain (goal, situation) pairs,
One may also define the notion of
then one can create intelligent systems
“efficient intelligence” [4] obtained by
displaying general intelligence with respect to
normalizing general intelligence according to
this simplicity measure, via subtly constructing
computational resources utilized. There is
the systems in a way that embodies the patterns
essentially no formal theory of efficient
implicit in the simplicity measure. From this
intelligence at this point. But efficient
perspective, there are two interesting approaches
intelligence is what matters in the real world,
to the “general theory of general intelligence
since theoretical general intelligences utilizing
under feasible computational resources”:
excessive amounts of computational resources

can never actually be constructed.
1. Discovering general principles that

According to this approach, different
determine how, given a simplicity
simplicity measures lead to different notions of
measure, to construct systems displaying
intelligence, and hence may correspond to
general intelligence relative to that
intelligent systems with very different
simplicity measure
properties. It seems extremely likely to us that,
2. Exploring a particular class of simplicity
if one measures simplicity as program length in
measures that seems to have high
some standard computational model (say, a
relevance, and figuring out how to create
classical Turing machine; or lambda calculus),
systems displaying general intelligence
then the best approach to general intelligence is
relative to simplicity measures in this
going to be some sort of simplistic brute force
class
search algorithm such as Hutter’s AIXI [5].

Essentially, what AIXI does, at each time point,
The former would be the more ambitious
is to search the space of all computer programs
approach: one can envision a theory that would
to find the program P so that, if AIXI uses P to
let one specify a simplicity measure, and would
2

---

then produce an AI design displaying a near-
and another agent is able to interpret the
maximal
amount
of
efficient
general
meaning
intelligence relative to that simplicity measure.
• Demonstrative communication, in
While we are interested in pursuing this
which an agent carries out a set of
direction, our focus in this paper is on the
actions in the world, and the other agent
second and less ambitious goal. In the
is able to imitate these actions, or
following section we will articulate a specific
instruct another agent as to how to
class of simplicity measures relevant to broadly
imitate these actions
humanlike intelligence, and then briefly explore
• Depictive communication, in which an
the properties of systems displaying general
agent creates some sort of (visual,
intelligence relative to these simplicity
auditory, etc.) construction to show
measures. A sister paper [6] then explores these
another agent, with a goal of causing the
implications more thoroughly, developing the
other agent to experience phenomena
notion of “cognitive synergy” that emerges from
similar to what they would experience
some of these properties as an architectural
upon experiencing some particular entity
principle for AGI design; and [7] then follows
in the shared environment
this up by looking at cognitive synergy in the
• Intentional communication, in which an
specific context of the OpenCogPrime AGI
agent explicitly communicates to another
architecture.
agent what its goal is in a certain

situation (note interesting recent results
2. THE EMBODIED COMMUNICATION
showing that mirror neurons fire in
PRIOR
response to some cases of intentional

communication as thus defined; [8])
Here we describe a simplicity measure on

(goal, environment) pairs; or, equivalently, a

It is clear that ordinary everyday
probability distribution over such pairs, which
communication between humans possesses all
we call the Embodied Communication Prior or
these aspects. The Embodied Communication
ECP. Here we describe the ECP intuitively
Prior is defined as the probability distribution in
rather than formalizing it rigorously; the latter
which the probability of an entity (e.g. a goal or
would be a feasible project but would result in a
environment) is proportional to the difficulty of
much longer and more technical paper.
describing that entity, for a typical member of
Consider a community of embodied
the community in question, using a particular
agents living in a shared world, and suppose that
set of communication mechanisms including the
the agents can communicate with each other via
above five modes. We will sometimes refer to
a set of mechanisms including:
the prior probability of an entity under this

distribution, as its "simplicity" under the
• Linguistic communication, in a
distribution.
language whose semantics is largely (not

Next, to further specialize the Embodied
necessarily wholly) interpretable based
Communication Prior, we will assume that for
on mapping linguistic utterances into
each of these modes of communication, there
finite combinations of entities drawn
are some aspects of the world that are much
from a finite vocabulary
more easily communicable using that mode than
• Indicative communication, in which e.g.
the other modes. For instance, in the human
one agent points to some part of the
everyday world:
world or delimits some interval of time,

3

---

• Abstract statements spanning large
current neurological or psychological data
classes of situations are generally much
currently supports.
easier to communicate linguistically

Of course there may be much
• Complex, multi-part procedures are
knowledge, of relevance to systems seeking
much easier to communicate either
intelligence according to the ECP, that does not
demonstratively, or using a combination
fall into any of these categories and constitutes
of demonstration with other modes
"mixed knowledge." There are some very
• Sensory or episodic data is often much
important specific subclasses of mixed
easier to communicate demonstratively
knowledge. For instance, episodic knowledge
• The current value of attending to some
(knowledge about specific real or hypothetical
portion of the shared environment is
sets of events) will most easily be
often much easier to communicate
communicated via a combination of declarative,
indicatively
sensory and (in some cases) procedural
• Information about what goals to follow
communication. Scientific and mathematical
in a certain situation is often much easier
knowledge are generally mixed knowledge, as is
to communicate intentionally, i.e. via
most everyday commonsense knowledge.
explicitly indicating what one’s own

Some cases of mixed knowledge are
goal is
reasonably well decomposable, in the sense that

they decompose into knowledge items that

These
simple
observations
have
individually fall into some specific knowledge
significant implications for the nature of the
type. For instance, an experimental chemistry
Embodied Communication Prior. For one thing
procedure may be much better communicable
they let us define multiple forms of knowledge:
procedurally, whereas an allied piece of

knowledge from theoretical chemistry may be
• Isolatedly declarative knowledge is that
much better communicable declaratively; but in
which is much more easily
order to fully communicate either the
communicable linguistically
experimental procedure or the abstract piece of
• Isolatedly procedural knowledge is that
knowledge, one may ultimately need to
which is much more easily
communicate both aspects.
communicable demonstratively

Also, even when the best way to
• Isolatedly sensory knowledge is that
communicate something is mixed-mode, it may
which is much more easily
be possible to identify one mode that poses the
communicable depictively
most important part of the communication. An
• Isolatedly attentive knowledge is that
example would be a chemistry experiment that
which is much more easily
is best communicated via a practical
communicable indicatively
demonstration together with a running narrative.
• Isolatedly intentional knowledge is that
It may be that the demonstration without the
which is much more easily
narrative would be vastly more valuable than
communicable intentionally
the narrative without the demonstration. To

cover such cases we may make less restrictive


definitions such as
This categorization of knowledge types

resembles many ideas from the cognitive theory
• Interactively declarative knowledge is
of memory [9, 10], although the distinctions
that which is much more easily
drawn here are a little crisper than those which
communicable in a manner dominated
by linguistic communication
4

---

3. LEVELS OF GENERAL
and so forth. We call these “interactive
INTELLIGENCE
knowledge categories,” by contrast to the

“isolated knowledge categories” introduced
Another, complementary way to explore the
earlier.
relationship between the ECP and real-world

Next we introduce an assumption we call
general intelligence, is to tie it in with various
NKC, for Naturalness of Knowledge Categories.
relevant theories of the levels of general
The NKC assumption states that the knowledge
intelligence.
in each of the above isolated and interactive

Piaget [11] introduced a series of
communication-modality-focused
categories
cognitive developmental stages, which may be
forms a "natural category," in the sense that for
approximately summarized as
each of these categories, there are many

different properties shared by a large percentage
1. Infantile: learns to carry out procedures
of the knowledge in the category, but not by a
oriented toward achieving goals in
large percentage of the knowledge in the other
relevant contexts
categories. This means that, for instance,
2. Operational: learns abstractions
procedural knowledge systematically (and
allowing it to adapt its learning to
statistically) has different characteristics than
different contexts and subgoals
the other kinds of knowledge.
3. Formal: learns abstractions allowing it

The
NKC
assumption
seems
to adapt its adaptation
commonsensically to hold true for human

everyday knowledge, and it has fairly dramatic
Piaget also introduced a “pre-operational”
implications for general intelligence. Suppose
stage between the infantile and operational ones;
we conceive general intelligence as the ability to
and some more recent thinkers have introduced
achieve goals in the environment shared by the
a “post-formal” stage involving the ability of the
communicating agents underlying the Embodied
system to apply formal reasoning to its own
Communication Prior. Then, NKC suggests
basic structure and outlook. But these three
that the best way to achieve general intelligence
phases will suffice for our current purposes.
according to the Embodied Communication

Taking more of a cybernetics and
Prior is going to involve
general systems theory approach, Gregory

Bateson [12] proposed a different sort of
• specialized methods for handling
hierarchy:
declarative, procedural, sensory and

attentional knowledge (due to the
1. Learning
naturalness of the isolated knowledge
2. Learning how to learn
categories)
3. Learning how to learn how to learn
• specialized methods for handling

interactions between different types of
Of course, there is no reason the hierarchy
knowledge, including methods focused
of levels of learning needs to stop at that point;
on the case where one type of
but Bateson suggests that in actual human
knowledge is primary and the others are
practice it generally does.
supporting (the latter due to the

In [13] it is suggested that these two
naturalness of the interactive knowledge
hierarchies, which look quite different on the
categories)
surface, are actually closely aligned, so that in
.
certain types of intelligent systems, the Piagetan

and Batesonian stages correspond closely.
5

---

Specifically it is argued that this correspondence
simplify analysis we consider an AI system that
holds both in humans and in AI systems whose
possesses a specific set of goals, together with a
operations are centrally based on uncertain
predefined set of actuators and sensors. We
inference. What we suggest here is that this
further assume that the action of the system may
correspondence is actually a conceptual
be modeled in terms of the “cognitive
consequence of the Embodied Communication
schematic”
Prior.


In the context of the ECP, the
Context & Procedure ? Goal <p>
operational stage may be viewed as requiring a

thorough integration of declarative and
interpreted to mean “If the context C appears to
procedural learning. Declarations about
hold currently, then if I enact the procedure P, I
procedures must be learned, and then
can expect to achieve the goal G with certainty
manipulated in order to lead to new procedures.
p.” A procedure is defined as some systematic
This is precisely Batesonian “learning how to
pattern of activity within the system (which may
learn.” Of course sensory knowledge must also
involve activation of the external actuators, or
be drawn into the picture here, as these
may in some cases involve purely internal
declarations will often be highly dependent on
activities). A context is defined as a fuzzy
sensorially-defined contexts.
logical predicate that holds, to a certain extent,

On the other hand, the formal stage
during each interval of time. A goal is simply
requires a yet deeper integration, in which
some fuzzy logical predicate that has a certain
procedures are learned for controlling the
value at each interval of time, as well. We will
application of declarative knowledge to guiding
also us the shorthand
procedure learning – and declarative knowledge

is then learned regarding these higher-level
C & P ? G <p>
procedures. This is a direct translation of the

ideas of [13] into the language of the ECP; and
Note that we don’t assume the system
clearly it is an instance of Batesonian “learning
explicitly uses the cognitive schematic to
how to learn how to learn.”
regulate its activities (nor uses fuzzy logical

Thus, we suggest that in the context of
predicates, or the other apparatus introduced
the ECP, Bateson’s third level of learning and
above, in its internal operations); rather, we are
Piaget’s formal stage come together in a
introducing this as an external model of the
dynamic of “mixed declarative/procedural
system. If the system explicitly uses some form
learning about declarative learning about
of the cognitive schematic in its internal
procedural learning.” And this dynamic, we
operations, that’s fine, but our analysis does not
propose, is critical to the achievement of a high
require this.
level of efficient general intelligence according
This
formalization
leads
to
a
to the ECP.
conceptualization of the internal action of an

intelligent system as involving two “key

learning processes”:
4. THE COGNITIVE SCHEMATIC


1. Estimating the probability p of a posited
Now we explore a little more deeply
C & P ? G relationship
what the above conclusions imply about the
2. Filling in one or two of the variables in
internal operations of AI systems displaying
the
cognitive
schematic,
given
general intelligence with respect to the ECP
assumptions regarding the remaining
(and adopting the NKC assumption as well). To
variables, and directed by the goal of
6

---

maximizing the probability of the
where the probability of C & P1 ? G is
cognitive schematic
known for some P1 related to P

• Isolatedly declarative or isolatedly
Again, we stress that we don’t assume the
sensory knowledge can be useful for
system’s internal dynamics are explicitly
estimating the probability of the
oriented around these two types of activity.
implication C & P ? G, in cases where
What we assume is that the system can be
the probability of C1 & P ? G is
modeled this way, basically as a combination of:
known for some C1 related to C

• Isolatedly declarative knowledge can be
1. Evaluating conjectured relationships
useful for estimating the probability of
between procedures, contexts and goals
the implication C & P ? G, in cases
(“analysis”)
where the probability of C & P ? G1 is
2. Conceiving novel possible relationships
known for some G1 related to G
between procedures, contexts and goals

(“synthesis”)

We can also see the role of mixed

knowledge here, because sometimes the best
Given this conceptualization, we can see
way to handle the schematic equation will be to
that, where synthesis is concerned,
fix only one of the terms. For instance, if we fix

G, sometimes the best approach will be to
• Isolatedly procedural knowledge can be
collectively learn C and P (for instance, using
useful for choosing P, given fixed C and
evolutionary learning methods, to allow them to
G
“co-evolve”).

Dominantly
procedural
• Isolatedly sensory knowledge, isolatedly
knowledge, for example, corresponds to the case
declarative knowledge, or mixed
where one mainly wants to learn P, but accepts
sensory/declarative knowledge can be
that one may also need to adapt C or G during
useful for choosing C, given fixed P and
the learning process, rather than leaving them
G
fixed.
• Isolatedly declarative knowledge can be

The final fact we need to account for is
useful for choosing G, given fixed P and
that, in any real-world context, a system will be
C
presented with a huge number of possibly

relevant analysis and synthesis problems.

On the other hand, where analysis is
Choosing which ones to explore is a difficult
concerned:
cognitive problem in itself – a problem that also

takes the form of the cognitive schematic, but
• Isolatedly declarative knowledge can be
where the procedures are internal rather than
useful for estimating the probability of
external. Thus this problem may be addressed
the implication in the schematic
via the analysis and synthesis methods describe
equation, given fixed C, P and G.
above. This is the role of attentional
Episodic knowledge can also be useful
knowledge; it gives the system some base
in this regard, via enabling estimation of
knowledge regarding what to attend to (which in
the probability via simple similarity
some cases will be the problem of using
matching against past experience.
complex analysis and/or synthesis to figure out
• Isolatedly procedural knowledge can be
what to attend to). Attentional knowledge may
useful for estimating the probability of
be built up analytically or synthetically, and
the implication C & P ? G, in cases
isolatedly or interactively, just like other types
of knowledge. The NKC suggests that
7

---

attentional knowledge forms a natural category
type K1 distinct from K, there should be
just like the other types of knowledge.
a distinctive capability with interaction

Suppose we conceive an AI system as
type “interactive” and dealing with
consisting of a set of learning capabilities, each
knowledge that is interactively K but
one characterized by three features:
also includes aspects of K1


• One or more knowledge types that it is

Furthermore, it seems intuitive that
competent to deal with, in the sense of
according to the ECP with NKC, if the
the two key learning problems
capabilities mentioned in the above points are
mentioned above
reasonably able, then the system possessing the
• At least one learning type: either
capabilities will display general intelligence
analysis, or synthesis, or both
relative to the ECP. Thus we arrive at the
• At least one interaction type, for each
hypothesis that
(knowledge type, learning type) pair it

handles: “isolated” (meaning it deals
Under the assumption of the Embodied
mainly with that knowledge type in
Communication Prior (with the Natural
isolation), or “interactive” (meaning it
Knowledge
Categories
assumption),
the
focuses on that knowledge type but in a
property above called “cognitive completeness”
way that explicitly incorporates other
is necessary and sufficient for efficient general
knowledge types into its process), or
intelligence at the Piagetan formal level.
“fully mixed” (meaning that when it

deals with the knowledge type in

Of course, the above considerations are
question, no particular knowledge type
very far from a rigorous mathematical proof (or
tends to dominate the learning process).
even precise formulation) of this hypothesis.

But we are presenting this here as a conceptual

Then, it seems to follow from the ECP
hypothesis, in order to qualitatively guide R&D
with NKC that systems with high efficient
and also to motivate further, more rigorous
general intelligence should have the following
theoretical work.
properties, which collectively I’ll call “cognitive

An approach to AGI architecture called
completeness”:
“Cognitive Synergy Theory” [6] goes into more

detail regarding the types of cognitive process
• For each (knowledge type, learning type,
involved in intelligent systems modeled by the
interaction type) triple, there should be a
cognitive schematic, and the ways in which they
learning capability corresponding to that
may interact with each other and support each
triple.
other. [7] then looks at the relationship of
• Furthermore the capabilities
Cognitive Synergy Theory with a specific AGI
corresponding to different (knowledge
architecture, OpenCogPrime.
type, interaction type) pairs should have

distinct characteristics (since according
5. BEYOND THE EMBODIED
to the NKC the isolated knowledge
COMMUNICATION PRIOR
corresponding to a knowledge type is a

natural category, as is the dominant

One interesting direction for further
knowledge corresponding to a
research would be to broaden the scope of the
knowledge type)
inquiry, in a manner suggested above: instead of
• For each (knowledge type, learning type)
just looking at the ECP, look at simplicity
pair (K,L), and each other knowledge
measures in general, and attack the question of
8

---

how a mind must be structured in order to
lacking true mathematical rigor, the ideas
display efficient general intelligence relative to
presented have been semi-formalized, in an
a specified simplicity measure. This problem
attempt to maintain a high level of conceptual
seems unapproachable in general, but some
precision; but nevertheless, they must be
special cases may be more tractable.
considered at this stage to overlap as much with

For instance, suppose one has
the category of “philosophy of mind” as with

mathematics or science.
• a simplicity measure that (like the ECP)

Yet, we do not consider these ideas
is approximately decomposable into a
pragmatically irrelevant due to their somewhat
set of fairly distinct components, plus
philosophical nature. Given the current lack of
their interactions
a rigorous mathematical and scientific theory of
• an assumption similar to NKC, which
real-world general intelligence, those of us
states that the entities displaying
concerned with constructing AGI systems or
simplicity according to each of the
analyzing human intelligence require some sort
distinct components, are roughly
of guidance. And, it seems better to draw
clustered together in entity-space
guidance from conceptually clear, only partially

rigorous thinking about the right issues, than

Then one should be able to say that, to
from more rigorously mathematically or
achieve efficient general intelligence relative to
empirically grounded thinking about the wrong
this decomposable simplicity measure, a system
issues .
should have

As the discussion here has hopefully

made clear, theorizing about “fully general
• distinct capabilities corresponding to
intelligence” is not likely to be useful for
each of the components of the simplicity
understanding real-world general intelligence.
measure
And yet, one would like to be able to say
• interactions between these capabilities,
something about general intelligence going
corresponding to the interaction terms in
beyond the description of particular systems like
the simplicity measure
human brains or particular software systems.

On the other hand, from an AGI point of view,

With copious additional work, these
studying the human brain and mind is valuable,
simple observations could potentially serve as
but can also be confusing and misleading, if
the seed for a novel sort of theory of general
one’s goal is not to precisely emulate human
intelligence – a theory of how the structure of a
intelligence but rather to make a different,
system depends on the structure of the
perhaps in some respects better, sort of
simplicity measure with which it achieves
intelligence, operating in the same everyday
efficient general intelligence. Cognitive
world as humans.
Synergy Theory would then emerge as a special

The approach taken here seeks to find a
case of this more abstract theory.
middle path, via qualitatively characterizing the

class of systems that are generally intelligent
6. CONCLUSION
with respect to the Embodied Communication

Prior (with the Natural Knowledge Categories

We have introduced a novel approach to
assumption); a class which, intuitively, appears
defining
real-world
general
intelligence,
to include both human brains, and the systems
attempting to occupy the middle ground
that would result from fully implementing and
between excessive mathematically generality
teaching certain contemporary AGI designs
and excessive anthropomorphism. While
such as OpenCogPrime.
9

---

One promising future research direction
[5] Hutter, Marcus (2005). Universal AI.
would be to attempt to create a fully
Springer.
mathematically rigorous version of the ideas
[6] Goertzel, Ben (2009). Cognitive Synergy: A
presented here – a quest that would doubtless
Universal Principle of Feasible General
involve a number of refinements and revisions
Intelligence?, Dynamical Psychology
of the ideas, though I suspect that the spirit
[7] Goertzel, Ben (2009a). OpenCog Prime: A
would remain intact. Another, noted above, is
Cognitive Synergy Based Architecture for
to extend the scope of these ideas more
Artificial General Intelligence. Dynamical
thoroughly beyond the ECP to deal with more
Psychology.
general classes of simplicity measures. One
[8] Fogassi, Leonardo, Pier Francesco Ferrari,
could also attempt to go further in the direction
Benno Gesierich, Stefano Rozzi,
of directly deriving AGI designs from these
Fabian Chersi, Giacomo Rizzolatti (2005).
theoretical considerations (rather than using the
Parietal lobe: from action organization
theory to analyze existing AGI designs, as done
to intention understanding. Science 308: 662-
in [7]), or using these considerations to analyze
667.
human brain function. All these avenues seem
[9] Eichenbaum (2002). The Cognitive
valuable, along with others; clearly the study of
Neuroscience of Memory. Oxford University
general intelligence is still at a very early stage,
Press.
theoretically as well as pragmatically.
[10] Tulving and Craik (2005). The Oxford

Handbook of Memory. Oxford University
References
Press.

[11] Piaget, Jean. (1955) The Construction of
[1] Legg, S. and Marcus Hutter. (2007) A
Reality in the Child. Routledge and Kegan
Collection of Definitions of Intelligence. In B.
Paul.
Goertzel, editor, Advances in Artificial General
[12] Bateson, Gregory (1980). Mind and
Intelligence, IOS Press
Nature: A Necessary Unity. Bantam
[2] Legg, S. and Marcus Hutter. (2006). A
[13] Goertzel and Bugaj (2007). Stages of
Formal Measure of Machine Intelligence. In
Cognitive Development in Uncertain Inference
Proc. Annual machine learning conference of
Based AI Systems. In B. Goertzel, editor,
Belgium and The Netherlands (Benelearn-
Advances in Artificial General Intelligence, IOS
2006). Ghent, 2006.
Press
[3] Goertzel, Ben (1993). The Structure of

Intelligence. Springer.

[4] Goertzel, Ben (2006). The Hidden Pattern.

BrownWalker.

10

---