Harnad, S. (1990) The Symbol Grounding Problem. Physica D 42: 335-346.
http://www.cogsci.soton.ac.uk/~harnad/Papers/Harnad/harnad90.sgproblem.html
The symbol grounding problem is a difficulty found with formal symbol
systems as a cognitive theory, ie. a possible explanation/illustration
of the workings of the human brain and mind. An alternative cognitive
theory being discussed (and relevant to Harnad's arguments later) is
that of connectionism or neural networks, so Harnad starts by briefly
explaining both symbol systems and neural nets...
> HARNAD:
> 1.2 Symbol Systems
> What is a symbol system? ... A symbol system is:
> 1. a set of arbitrary "physical tokens" scratches on paper, holes on a tape,
> events in a digital computer, etc. that are
> 2. manipulated on the basis of "explicit rules" that are
> 3. likewise physical tokens and strings of tokens. The rule-governed symbol-
> token manipulation is based
> 4. purely on the shape of the symbol tokens (not their "meaning"), i.e., it is
> purely syntactic, and consists of
> 5. "rulefully combining" and recombining symbol tokens. There are
> 6. primitive atomic symbol tokens and
> 7. composite symbol-token strings. The entire system and all its parts -- the
> atomic tokens, the composite tokens, the syntactic manipulations both actual
> and possible and the rules -- are all
> 8. "semantically interpretable:" The syntax can be systematically assigned a
> meaning e.g., as standing for objects, as describing states of affairs).
This is an unsurprisingly similar model to the one we have been
discussing in lectures, apart from the omission of the 'implementation
independent' bit, which Harnad explains less emphatically in the next
paragraph...
> HARNAD:
> Symbolists emphasize that the symbolic level (for them, the mental level) is a
> natural functional level of its own, with ruleful regularities that are
> independent of their specific physical realizations. For symbolists, this
> implementation-independence is the critical difference between cognitive
> phenomena and ordinary physical phenomena and their respective explanations.
Now neural nets...
> HARNAD:
> 1.3 Connectionist systems
> Variously described as "neural networks," "parallel distributed processing"
> and "connectionism," this approach has a multiple agenda, which includes
> providing a theory of brain function...
> Connectionism will .. only be considered here as a cognitive theory.
> As such, it has lately challenged the symbolic approach to modeling the mind.
> According to connectionism, cognition is not symbol manipulation but dynamic
> patterns of activity in a multilayered network of nodes or units with weighted
> positive and negative interconnections. The patterns change according to
> internal network constraints governing how the activations and connection
> strengths are adjusted on the basis of new inputs (e.g., the generalized
> "delta rule," or "backpropagation," McClelland, Rumelhart et al. 1986). The
> result is a system that learns, recognizes patterns, solves problems, and
> can even exhibit motor skills.
> 1.4 Scope and Limits of Symbols and Nets
> It is far from clear what the actual capabilities and limitations of either
> symbolic AI or connectionism are. The former seems better at formal and
> language-like tasks, the latter at sensory, motor and learning tasks, but
> there is considerable overlap and neither has gone much beyond the stage of
> "toy" tasks toward lifesize behavioral capacity.
Harnad doesn't actually make clear here what he would consider a 'lifesize'
task.
> HARNAD:
> Moreover, there has been some
> disagreement as to whether or not connectionism itself is symbolic. We will
> adopt the position here that it is not, because connectionist networks fail to
> meet several of the criteria for being symbol systems, as Fodor & Pylyshyn
> (1988) have argued recently. In particular, although, like everything else,
> their behavior and internal states can be given isolated semantic
> interpretations, nets fail to meet the compositeness (7) and systematicity (8)
> criteria listed earlier: The patterns of interconnections do not decompose,
> combine and recombine according to a formal syntax that can be given a
> systematic semantic interpretation.[5] Instead, nets seem to do what they do
> non symbolically.
I think that I agree with this view intuitively, as well as because of
the logical reasons given here. The more 'natural' process of training
a network, and the evolution of weightings which can even produce
time-dependent outputs where required, seems fundamentally different to
formal symbol systems. The question really comes up because neural nets
are often simulated by symbol systems (ie. your average digital
computer) instead of being fully implemented, but Harnad explains this
quite well in his footnote...
> HARNAD:
> [5.] There is some misunderstanding of this point because it is often
> conflated with a mere implementational issue: Connectionist networks can be
> simulated using symbol systems, and symbol systems can be implemented using a
> connectionist architecture, but that is independent of the question of what
> each can do qua symbol system or connectionist network, respectively. By way
> of analogy, silicon can be used to build a computer, and a computer can
> simulate the properties of silicon, but the functional properties of silicon
> are not those of computation, and the functional properties of computation are
> not those of silicon.
Harnad then introduces a motivation for his presentation of the symbol
grounding problem, namely that although connectionism has been
criticised for not being symbolic and so not representative of the more
logical aspects of the mind, yet symbolism is not without its own
problems.
Harnad begins by describing Searle's Chinese Room argument as an
example of the symbol grounding problem, which Searle called
intentionality or intrinsic meaning. I will skip the explanation of the
argument as we all know it so well now, but here is Harnad's
explanation of its relevance to the problem in question...
> HARNAD:
> The symbols and the symbol manipulation, being all based on shape rather than
> meaning, are systematically interpretable as having meaning -- that, after
> all, is what it is to be a symbol system, according to our definition. But the
> interpretation will not be intrinsic to the symbol system itself: It will be
> parasitic on the fact that the symbols have meaning for us, in exactly the
> same way that the meanings of the symbols in a book are not intrinsic, but
> derive from the meanings in our heads. Hence, if the meanings of symbols in a
> symbol system are extrinsic, rather than intrinsic like the meanings in our
> heads, then they are not a viable model for the meanings in our heads:
> Cognition cannot be just symbol manipulation.
As an aside, it may or may not be worth mentioning that I'm not
actually convinced of the efficacy of Searle's argument. I'm not saying
that I think the 'program' or 'computer' described does have a mind or
even an understanding of Chinese (though it's possible that depends on
your understanding of 'an understanding'!), in fact I would tend
towards saying it didn't. I simply don't think that Searle makes a
particularly relevant proof (and no, I can't think of a better). I'll
try to explain myself - when Searle is looking at the Chinese symbols
and generating his replies, he is presumably using (a) different
part(s) of his brain than if he had been reading and writing English
(ie. step-by-step following of memorised rules, rather than the more
sublimated absorption of recognised language). Therefore it cannot be
expected that Searle himself will have any understanding of the Chinese
(which is presumably the conclusion Searle expects us to draw).
However, to me this implies that Searle is asking the wrong question -
it is not the language recognition part of his brain which is analysing
the Chinese, so his 'mind' (whether we are talking about another
physical part of the brain or some pseudo-physical amalgamation greater
than the brain) does not understand it, but why does that necessarily
mean that the language recognition part of the computer/program does
not generate some kind of understanding 'elsewhere'?
That aside, the Chinese Room argument does give a good example of the
symbol grounding problem, and I think this paper lends importance to
Searle's discussion, whether you originally agreed with it or not.
Harnad then gives another example we have briefly discussed, the
Chinese/Chinese dictionary.
> HARNAD:
> This is more like the actual task faced by a
> purely symbolic model of the mind: How can you ever get off the symbol/symbol
> merry-go-round? How is symbol meaning to be grounded in something other than
> just more meaningless symbols?[9] This is the symbol grounding problem.[10]
OK, so having finally put his finger on what the problem is, Harnad
begins describing his proposed solution...
> HARNAD:
> a hybrid nonsymbolic/symbolic system, a "dedicated" one, in which the
> elementary symbols are grounded in two kinds of nonsymbolic representations
> that pick out, from their proximal sensory projections, the distal object
> categories to which the elementary symbols refer.
> [Connectionism and Symbolism] respective strengths will be put to cooperative
> rather than competing use in our hybrid model, thereby also remedying some of
> their respective weaknesses.
To justify this system and to give some of its aims, Harnad gives a
model of human 'behavioral capacity'...
> HARNAD:
> They can (1) discriminate, (2) manipulate, (3) identify and (4) describe
> the objects, events and states of affairs in the world they live in, and they
> can also (5) "produce descriptions" and (6) "respond to descriptions" of those
> objects, events and states of affairs...
> To discriminate is to able to judge whether two inputs are the same or
> different, and, if different, how different they are.
> To identify is to be able to assign a unique (usually arbitrary) response - a
> "name" - to a class of inputs, treating them all as equivalent or invariant in
> some respect.
Harnad argues that to be able to discriminate and identify objects and
classes, we need an internal representation, and proposes 'iconic
representations', analog transforms of received sensory input. It would
then be a relatively simple process to measure two icons degree of
similarity. However, as Harnad discusses, icons are not sufficient for
identification as there would simply be too many of them, and their
distinctions too vague. This is similar to the fictional character we
were discussing on Thursday, who remembered everything so was unable to
classify objects because he remembered them all (and all instances of
each) as distinct. Instead, Harnad says, the unimportant must be
forgotten...
> HARNAD:
> For identification, icons must be selectively reduced to those "invariant
> features" of the sensory projection that will reliably distinguish a member of
> a category from any nonmembers with which it could be confused. Let us call
> the output of this category-specific feature detector the "categorical
> representation" ...
> Note that both iconic and categorical representations are nonsymbolic. The
> former are analog copies of the sensory projection, preserving its "shape"
> faithfully; the latter are icons that have been selectively filtered to
> preserve only some of the features of the shape of the sensory projection:
> those that reliably distinguish members from nonmembers of a category. But
> both representations are still sensory and nonsymbolic. There is no problem
> ... of semantic interpretation, or whether the semantic interpretation is
> justified.
OK, so we now have these icon thingies, such that if the system sees a
horse, it can say 'Horse!', but very little else - we have a
classifier. And as Harnad discusses a little later, a very good
candidate for classifiers is connectionism, ie. neural nets. I don't
know how many of the class did Neural Nets last semester, but this is a
classic application for them - apply a feature map as input, and the
net generates an output indicating one particular class. Very nice,
but we now want to do something with our 'Horse!'...
> HARNAD:
> For systematicity it must be possible to combine and recombine [categorical
> representations] rulefully into propositions that can be semantically
> interpreted. What would be required to generate these other systematic
> properties? Merely that the grounded names in the category taxonomy be
> strung together into propositions about further category membership
> relations.
Harnad gives the example that, given awareness of the icons for 'horse'
and 'stripes', it should be a simple matter to define a new concept
'zebra' as the conjunction of these two, and this concept would
'inherit' a grounding from them.
> HARNAD:
> Once one has the grounded set of elementary symbols provided by a taxonomy of
> names (and the iconic and categorical representations that give content to the
> names and allow them to pick out the objects they identify), the rest of the
> symbol strings of a natural language can be generated by symbol composition
> alone,[18] and they will all inherit the intrinsic grounding of the elementary
> set.[19] Hence, the ability to discriminate and categorize (and its underlying
> nonsymbolic representations) has led naturally to the ability to describe and
> to produce and respond to descriptions through symbolic representations...
> 5. Conclusions
> ... If both tests are passed [the 8 criteria for a symbol system, and the
> behavioral stuff mentioned earlier], then the semantic interpretation of its
> symbols is "fixed" by the behavioral capacity of the dedicated symbol system,
> as exercised on the objects and states of affairs in the world to which its
> symbols refer; the symbol meanings are accordingly not just parasitic on the
> meanings in the head of the interpreter, but intrinsic to the dedicated symbol
> system itself. This is still no guarantee that our model has captured
> subjective meaning, of course. But if the system's behavioral capacities are
> lifesize, it's as close as we can ever hope to get.
It may be a boring answer, but I don't have any real problems with this
paper. I agree that the symbol grounding problem is one intrinsic to
symbolic AI, and I like the simplicity of the idea behind Harnad's
solution. I think a good implementation would be a very interesting
experiment, whether you are a believer in weak or strong AI.
Butterworth, Penny <pjb297@ecs.soton.ac.uk>
This archive was generated by hypermail 2b30 : Tue Feb 13 2001 - 16:36:27 GMT