Poetry: Shakespearean sonnets are in iambic pentameter and consists of three quatrains followed by a couplet»
[ACG:3f].
«That the computer ... has manlike properties should not surprise us.
It was invented to do intellectual work for us. Thus, we are naturally
antropomorhic in our description and understanding of computers.
In particular, we expect to carry on an extensive symbolic dialogue
with our computers in languages acceptable to us (problem languages).
...
Each computer has its own control language, usually quite remote
from the wide variety of languages demanded by the diverse uses of the machine.
A translation from problem language to machine language is required.
Assemblers, interpeters, and compilers have been developed to transfer as much as possible
of this tedious task from man to machines» [ACG:4f]
There are five language levels w.r.t. to symbolizing/representing/expressing knowledge [FO 632f]
|
level
| primitives | interpretation | main feature |
|
logical | predicates, functions | arbitrary | formalization |
|
epistemological | structuring relations | arbitrary | structure |
|
ontological | ontological relations | constrained | meaning |
|
conceptual | conceptual relations | subjective = psychological? | conceptualization |
|
linguistic | linguistic terms | subjective = cultural? | language dependency
|
- «At the (first order) logical level, the basic [language] primitives
are predicates and functions, which have given a formal semantics in terms
of relations among objects of a domain. No particular assumption is made
however on the nature of such relations, which are completely general and
content-independent. The logical level is the level of formalization:
it allowsfor a formal interpretation of the primitives, but their
interpretation remains totally arbitrary.»
For example, at the logical level `a red ball' could be represented
as 
x. Ball(x) & Red(x).
- At the epistemological level «the [language] primitives allow
to specify "the formal structure of conceptual units and their
interrelationships as conceptual units (independeent of any
knowledge expressed therein)". In other wordls, while the logical level
deals with abstract predicates and the conceptual level with specific
concept, at the epistemolofical level the generic notion of a
concept is introduced as a knowledge structuring primitive.
Concepts themselves--which correspond to unary predicates at the logical
level--have an internal structure, as they "bundle" together further
concepts or binary relations (roles). The epsitemological level is
tehrefore the level of structure.»
For example, it could be declared that in x. Ball(x) & Red(x)
(logical level), Ball is a concept and
Red is a filler of a Colour role.
- «At the ontolological level, such ontological commitments
associatied to the language primitives are specified explicitly. Such a
specification can be made in two ways: either by suitably restricting the
semantics of the primitives, or by introducing meaning postulates expressed
in the language itself. In both cases, the goal is to restrict the number
of possible interpretatoins, characteriying the meaning of the basic
ontological categories used to describe the domain: the ontological level
is therefore the level of meaning. Of course such a
characterization will be in general incomplete, and the result will be an
approximation of the set of intended models. ...»
For example, it could be fixed what concepts, roles and role-fillers are,
rendering it impossible to declare (at the epistemological level)
Red not as a role-filler but as a concept in accordance with the
sense of "red" we have in mind.
- «At the conceptual level, primitives have a definite cognitive
interpretation, corresponding to language-independent concepts like
elementary actions or themaic roles. The skeleton of the domain structure
is already given, independently of an explicit account of the underlying
ontological assumptions. Within a certain application domain, the user is
forced to express knowledge in the form of a specialization of this
skeleton. ...»
For example, a set of concepts and roles can be pre-defined
like PhysicalBody and Colour,
if they are agreed to be standard to a domain.
«However, our chances of getting such an agreement ... [depends] in thic
case on the principles we have adopted for the definition of our
basic ontological categories ... Notice that the necessity of well-founded
principles is much more relevant if we want to further specialize logical
relations into categories like parts, qualities, properties, states and so
on.»
- «Finally, primitives at the linguistic level directly refer to
verbs and nouns.»
Ontological distinctions
«allow the knowledge engineer to make clear the intended meaning
of a particular symbol. This is especially important since we are
constantly using natural language words for predicate symbols, relying on
them to make our statements readable and to convey meanings not explicitly
stated. However, since words are ambiguous in natural language, it may be
important to "tag" these words with a semantic category, in assoication
with a suitable axiomatisation, in order to guarantee a consistent
interpretation» [FO 636].
moved to T.W.O.
[This section is here until it finds a better place]
In (ultimate) system theory, meaning of signals/signs
in an information system is treated as follows:
Take 1. A system (core) S may react to a situation a by action z
through a control chain a -> ... -> x -> ... -> z.
The intermediate, system-internal signal x can be considered a sign
with two meanings to S (sign interpretant):
- the situation a ("cognitive meaning"), and
- the action z ("intentional meaning") [SuB 328].
(This is in the ideal case - in reality,
preceived sitation a may deviate from actual situation a', and
intended behavior z may deviate from actual behavior z'.)
The semantic quality which a sign x, like pointer angle, has
is the organetic quality of the designate a (e.g., car speed),
or z, respectively.
Take 2.
Any signal x is semanticable (potential sign?)
if it is in the system "core" S (sign interpretant)
and is "relevant" to S, ie., affects its fitness.
Now let x be a semanticable sign signal
which is conductor of a "homoeostatic dyad" <a;z> (see below)
in an idealized model.
The meaning of x is a semantically encoded description of a or z
wrt. a's interference with z's homoeostatis.
- x's cognitive meaning is the description of a
("valence" of a wrt. z)
- x's intentional meaning is the description of z
("problem" of z wrt. a) [SuB 359].
Hence meaning is a function of the system's structure [SuB 361].
A homoeostatic dyad <a;z>
is a noisefree homoeostatic subsystem with source a and sink z [SuB 338].
There are three type K(1) control blocks
(subsystems) which can establish homoestatic dyads [SuB 335]
(shown in form of Mason diagrams):
| homoeostatic mesh | homoeostatic cycle
(negative feedback loop) | homoeostatic net
|
.---------------v
a ----> z x
^-------'
a ----> z ----> x
^-------'
.---------------v
a ----> z ----> x
^-------'
|
a is the free input signal, the source of the homoeostatic entropy.
z is the "homoeostatic variable",
which depends on a and x in a way
that their possible changes are neutralized
(the sink of the homeostatic entropy).
x is the "conductor" of the homoeostatic entropy,
i.e., depends on a and affects z.
| | | |
|
A technical example [OOO 206]: a file cache is always consistent
if all reads and write to the cache go through to the disk,
ie. an effective coupling (but then there were no caching).
A cache optimizes this by not reading and writing the content of the file
but some complex but fast coordination mechanism.
The purpose of caching is that the file can be obtained from the cache
without needing to access the disk
(except at those time it knows the file's contents on the disk has changed,
then it does resynchronize with the disk).
IOW a cache continues to track the contents of a file on the disk
without being all the time effectively coupled to the disk.
|
Sunflowers rotate their head to follow the sun,
and, let's assume, stand still when the sun is hidden (behind a tree, house, etc.).
The super-sunflower continues rotating even while the sun is hidden.
«What distinguishes super-sunflowers is not the fact that they track the sun.
... [T]hey track something to which they are not effectively coupled.
This behavior, which I will call "non-effective coupling,"
is no less than the forerunner of semantics:
a very simple form of effect-transcending coordination in some way essential
to the overall existence or well-being of the constituted system» [OOO 203].
«There is nothing more basic to intentionality
than this pattern of coming together and coming apart,
at one moment being fully engaged, at another point being separated,
but separated ... in a way as to stay corrdinated with what
at the moment is distal and beyond effective reach» [OOO 206].
«[T]his coordination will in general be approximate. ...
[C]oordination can only settle on some but not all aspects
of the distal sitatuation» [OOO 207].
«[T]he super-sunflower, when it can no longer be driven by the
incident radiation, has to employ a different, internal mechanism ...
in order to continue to "track" the sun.
This retractionof responsibility into the s-region, as I will call it---this
shouldering of effective responsibility by the s-region,
to compensate for the break in effective coupling---is not less
than the origin of reasoning, representation and syntax:
effective projection, onto the intentional agent, of the requisite
arrangements for maintaining long-distance (semantic) coherence» [OOO 221].
«[R]esponsibility for mainining coordination across the gap
is asymmetrical. It "falls to the s-region," one might say---except
that gets it backwards. Rather: to be an s-region is to assume this
responsibility. ...
[T]he "distance" over which the s-region is able to maintain coordination
will depend on its cleverness---not just its internal cleverness,
but the total sophistication of the environment and cultural resources
on which it can lay its hands» [OOO 223f].
About neural network models:
«One area that is disturbing is that neither the network nor the programmer
have easy access to the information structures stored in the weights
of the hidden layer. For example, if the network learns to distinguish
the sonar echoes of different types of rocks and mines,
the information for all the different kinds of rocks and mines in the training
set will be superposed across the entire network. There is no easy way
to extract from the network the features of the mines that enable them
to be distinguished. ...
[E]ven though the network can clearly distinguish the mines from the rocks,
there is an important sense in which the network doesn't 'know' anything
about the usual or unusual shapes for rocks or mines.
Cussins (1990) offers an interesting distinction between
the conceputal and nonconceptual content
of a system that may help to clarify the problem. ...
[O]ur rock and mine connectionist system has the property of being able
to distinguish rocks and mines, but the system itself need not have the
concepts of rocks and mines in order to have the property to distinguish them.
Such a system 'knows its way around' a domain, but lacks the concepts
to describe what it knows. First-order connectionist networks are
ideal contenders for systems capable of supporting nonconceptual content.
The problem with nonceptual content is that there is no access
to the information that enables the system to negotiate a domain.
... [On the other hand it] seems that part of the ability to adapt to change
is to isolate and manipulate information and knowledge.
Andy Clark (1993) discusses Cussins' criteria ...:
"Cussins offers a rather intricate account in which one of the leading ideas
is that to have a conceptual content requires more than a mere
causal or informational link to the state of the world implicated
in the description of the content. To have (properly) the concept 'fly'
involves more than being able to find your way around (like a frog)
in a fly-infested domain. It involves having a whole web of concepts
in which your concept of a fly is embedded. In particular, it involves
having your fly concept at the disposal of any other conceptual
abilities you have."
This description of conceptual content sound very like the
"knowledge-level" approach in traditional AI.»
[CGPPF]
Connectionist networks and knowledge representation:
«There is a growing trend in the connectionist literature to reject the kind
of ontological engineering required in representing large amounts of
knowledge for traditional AI projects. ...
However, despite the power of connectionist systems to model learning and
categorization, I blieve that it is a mistake to reject outright
the "knowledge-level" approach. ...
... Some analysis is required in determining the architecture of the network,
the selection and encoding of the input data, and the category that each
data entry represents. But the actual set of weights in the network
that 'represents' the different concepts are not pre-determined.
Clearly in a connectionist system the level of representation is finer-grained
than one at the knowledge-level. For this reason, connectionist networks
are often said to work at the sub-symbolic or sub-linguistic level
and to use micro-features in categorization.
Connectionist systems are massively parallel networks where information
is stored in parameters associated with connections rather than in the
elements themselves. Thus, connectionist networks do not involve computations
defined over symbols. ... Basically, neural networks are good at many of the
tasks traditional AI programs are so bad at:
pattern recognition, learning and generalization» [CGPPF].
Ulf Schünemann 201202
Knowing
)
or a whole formal mathematical system (tensor calculus);
or we may assign new meanings to old forms (what does "cool" mean?).
There are examples of invented languages in law and differential calculus,
knitting and matrix algebra, mathematical logic and organic chemistry.
In each case the invented language has special symbolic properties
that simplify communication and comprehension.» [ACG:2]