Experience and
Sensation
Sellars and Dewey on
the Non-cognitive aspects of
Mental Life.
Published in "Education and Culture: the Journal of the John Dewey Society" Winter, 2001
Abstract
Sellars
and Dewey each isolated and critiqued different aspects of the atomistic epistemology
of the logical positivists: Dewey
labeled his target "Sensationalistic Empiricism", and Sellars labeled
his "the Myth of the Given."
The main theme of this paper will be the similarity and differences in
their responses to this kind of philosophy, and how both responses can be
clarified and strengthened by considering recent discoveries in Cognitive
Neuroscience. What we have recently learned about neural architecture accounts for a distinction between
knowledge and experience that is a recurrent theme in both Sellars and Dewey.
Dewey, however, made a sharper break from the positivists by seeing all
experience as shaped by skills and abilities which were designed to acheive
certain goals and were colored by emotions. The connectionist architecture used
in Cognitive Neuroscience supports this view, as does the psychological
research of J.J. Gibson. Once we consider the ways in which connectionist
cognitive abilities differ from linguistic ones, Sellars' distinction between
thoughts and sensations, and Dewey's distinction between knowledge and
experience, can both be plausibly accounted for.
Sellars
never used the word 'experience' as a technical term the way Dewey did, and to my knowledge never cited Dewey's
epistemology when discussing his own. But they were both interested in coming
up with an holistic alternative to the atomistic epistemology of Bertrand Russell,
G.E. Moore, and C.D. Broad.
Sellars and Dewey each isolated and critiqued different aspects of this
atomistic epistemology by giving it a name: Dewey labeled his target
Sensationalistic Empiricism, and Sellars called the object of his critique the
Myth of the Given. The main theme
of this paper will be the similarity and differences in their responses to this
kind of philosophy, and how both responses can be clarified and strengthened by
considering recent discoveries in Cognitive Neuroscience.
Shortly
after Dewey's death, it appeared that his critiques of Sensationalistic
Empiricism had been unsuccessful (at least in the academic marketplace). The
twentieth century empiricism that called itself "analytical philosophy" eventually replaced
Deweyan pragmatism in America, and various related theories of knowledge with
names like "Sense datum theory", "Logical Positivism," and
"Logical Empiricism" made an orthodoxy out of the view of knowledge
that Dewey had tried so hard to defeat. Now, however, this kind of empiricism
is in disrepute, thanks to critiques made primarily by Quine and Sellars, of
what Sellars called the Myth of the Given. Sellars' critique of sense datum
theory in his "Empiricism and
the Philosophy of Mind", and Quine's critique of the empiricist dogma
that it is possible to reduce all
meaningful statements to inferences from directly given observation sentences,
has made a new orthodoxy out of the belief that sensationalistic empiricism is
a dead horse no longer worth beating.
But
does the defeat of sensationalistic empiricism imply a triumph for the
alternative theory of experience that Dewey proposed in works like
"Experience and Nature"? Quine, at least, speaks favorably of Deweyan
pragmatism as the wave of the future, as do Rorty, Putnam and many others. But
Rorty seems to express the modern attitude when he says that "the
philosophers of today . . . tend
to talk about sentences a lot but say very little about ideas or
experiences" (Rorty 1994 p. 55). Quine certainly exemplifies this attitude
when half way through "two dogmas of empiricism", he decides to
sidestep the questions of what an observation is, and concentrate instead on
observation sentences. The fact that he proudly described himself as a
behaviorist, and spent a great deal of time attacking anything that might be
describable as mental, makes it highly unlikely that he would be sympathetic to
any talk about something called "experience". But Sellars, although
he refers to himself as a " verbal behaviorist", and considers
knowledge to be primarily linguistic, has a fair amount to say about the
non-linguistic aspects of experience. Unlike Quine, he emphasized that it was
necessary to talk about inner events, not just behavior and language.
Sellars and Sensations
Sellars
claimed that there were inner events he called "sensations" or
"sensa" , which seems to suggest a close kinship with sense data. But
he also avoided many of the mistakes of sense datum theory by turning the
concept of sensation on its head, making it a theoretical concept rather than
something immediately given. Once the supposedly essential characteristic of
givenness is removed from sensations, what is it that we have left, exactly?
There are many indications that Sellars sometimes thought that sense data do
exist, with almost all of the attributes that Broad, Moore and Russell ascribe
to them except for givenness.
I am sketching a view which
hovers on the edge of the sense datum theory, and yet I believe it succeeds in
avoiding falling into the abyss. (Sellars 1989 p.103)
For Sellars, sensations are
non-cognitive, because having a sensation is distinct from knowing about it.
They are also non-linguistic, because thinking "there is a pink ice
cube" is phenomenologically different from sensing a pink ice cube.
Sensations are described as "self-presenting' (ibid. p. 282), even though they must be accompanied by a
cognitive mental event for us to be aware that they are presenting themselves.[1] And once we do know about them that knowing is
considered to be "non-inferential". Unfortunately all of these descriptions are largely
negative, and although there are many attempts by Sellars to describe this kind
of knowing in positive terms, they are considered by many to be "one of
the most difficult and controversial aspects of his philosophy." (Rosenberg
1997)
Sellars,
like most Analytic and
Post-Analytic Philosophers, embraced the assumption that questions about
knowledge and thought are primarily questions about language. But unlike most
of his contemporaries, he was also aware of the limitations of this assumption.
In the following quotation, Sellars almost admits that there is something wrong
with his approach, without actually suggesting that there is any other way of
dealing with these questions.
Not all 'organized behavior' is
built on linguistic structures. The most
that can be claimed is that what
might be called 'conceptual thinking'
is essentially tied to language,
and that, for obvious reasons, the
central or core concept of what
thinking is pertains to conceptual
thinking. Thus, our common-sense
understanding of what sub-conceptual thinking -- e.g., that of babies and
animals -- consists in, involves viewing them as engaged in 'rudimentary' forms
of conceptual thinking. We
interpret their behavior using conceptual thinking as a model but qualify this model in ad hoc and
unsystematic ways which really amounts
to the introduction of a new notion which is nevertheless labeled
thinking'. Such analogical extensions of concepts, when supported by
experience, are by no means illegitimate. Indeed, it is essential to science.
It is only when the negative analogies are overlooked that the danger of
serious confusion and misunderstanding arises. (Sellars 1975 p.305)
Another
quotation, however, actually points towards another approach which was
eventually embraced by many of his students.
"the notion of sense data or
sensation itself is really a part of a scientific theory of perception, not a
philosophical theory. . . Philosophers often have to rush in where behaviorists
fear to tread. " (Sellars 1989 p. 119 ).
Since
Sellars said those words,
psychologists, neuroscientists and philosophers have all rushed in where
behaviorists had feared to tread, for behaviorism (and it's fear of opening the
black box that underlies human behavior and experience) is now as dead in psychology as sense datum theory is in
philosophy. And this willingness to inquire into the mechanisms underlying
thought has naturalized philosophy to a degree that often blurs the line
between philosophy and science. One of Sellars' students who leads this
naturalized philosophy movement is Ruth Millikan, who appears to have taken
both of the above paragraphs to heart in Millikan 1984. Instead of trying to
understand animal abilities by simply making analogies with human language-centered
thought, she formulated a whole new theory that saw both animal and human
cognition as members of more fundamental biological categories. The bare beginnings of that theory seem
to be prophesied in Sellars 1981, which even uses Millikan's paradigmatic bee
dances as an example of an animal
representational system. Another
important Sellars student, Paul Churchland, started a philosophical movement by extensively analyzing
the philosophical implications of the newest neurological data, and defending
the claim that because thought is fundamentally neurological, to conceive of it
as linguistic is at best misleading and worse false. This is a possibility that
was left open in Sellars 1981 when he said that although language is "our
primary concept of thinking, and therefore our conceptual point of entry into
our domain of the mental" it
was not primary in the order of being (p. 327). In other words, a careful study of neuroscience could
reveal that the ultimate nature of thoughts is as different from language as
tables and chairs are different from electrons. And it would not be surprising
that sensations, which are the big puzzle in a linguistically based theory of
mind, would be more comprehensible in a neurologically based one.
We
can thus see that there is good reason to think that a study of modern
Cognitive Neuroscience could be helpful in clarifying Sellars' epistemology. Is
there any reason to think this would also be true of Dewey's epistemology as
well?
Dewey's attitude towards epistemology
According
to Richard Rorty (whose views are in danger of becoming the orthodox modern
interpretation of Pragmatism[2]) the pragmatists at their best scorned
epistemology, and except during
confused moments of weakness like Dewey's "Experience and
Nature", believed it should be completely ignored . This
interpretation was probably
inspired by statements like this from Dewey.
. . . .intellectual progress
usually occurs through sheer abandonment of questions together with both of the
alternatives they assume--an abandonment that results from their decreasing
vitality and a change of urgent interest. We do not solve them: we get over
them. Old questions are solved by disappearing, evaporating, while new
questions corresponding to the changed attitude of endeavor and preference take
their place. (Dewey 1910, p. 19)
I
think, however, that Rorty goes further than Dewey intended
when he talks about disposing of epistemology by simply "changing the subject." For eliminating
questions about a subject is not the same as ignoring the subject
altogether. Astronomy did not
close up shop when it stopped asking the question "what turns the crystal
spheres?", and the fact that Dewey wants to abandon the traditional
epistemological questions does not mean he wants to abandon epistemology. He
does want to abandon epistemology's
traditional aspiration to be the most fundamental human discipline. For
he does not consider human beings to be fundamentally creatures that know, nor
does he assume that all other human activities can be understood by studying
how we think. Epistemology thus loses the position that Kant claimed for it
i.e. of being the foundation for all other disciplines. But the claim that the
subject of epistemology is not as fundamental as was once thought does not
imply that the subject itself disappears.
For
Dewey to study knowledge is to study inquiry, and inquiry is only one of many
human activities. It is a very important activity, perhaps the activity that is
most uniquely human. But knowing how to conduct inquiry is only one kind of
"knowing-how", that is, only one way that people relate to their
experience, and to each other in society.
Dewey's concept of experience is thus radically different from the
British Empiricists, who saw the flux of our daily experience to be a series of
confrontations with a succession of simple ideas, each confrontation resulting
in a kind of knowledge. For the British Empiricists, to be conscious was to be
aware, and to be aware was to possesses knowledge of what you are aware
of. Dewey rejected this, claiming instead that we have
experience prior to knowledge. Experience is what knowledge starts from, and is
its ultimate arbiter. If our theories cannot make sense out of our experience,
they must be changed. But experience itself is not knowledge; we get knowledge
only after inquiring into our experience.
This
hasty description of the relationship between knowledge and experience still
presupposes many ideas that Dewey
wants to reject. Deweyan
epistemology wants to shift from studying knowledge, which seems to be a static
"thing" that we acquire, to studying inquiry, which is an
activity. For Dewey,
"knowledge" was rather like a nominalization of a verb, rather than a
genuine noun, and he was vigilant in exposing the confusions that resulted from
ignoring this. Much of his "Democracy and Education" is devoted to
exploring the damage done by the assumption that knowledge is a commodity, and
that the goal of education is to cram as much of that commodity as possible
into a person's memory. ( A view of education that is still presupposed by many
practitioners of computational "GOFAI" artificial intelligence, who
assume that the way to create a more knowledgeable machine is to fill it with
the right sort of information, and teach it how to access that information
correctly.) When Dewey maps out
his five steps in the process of inquiry, it is clear that for him inquiry is a
specialized activity that is performed only at certain times, i.e. when things
don't work. When things do work, and we feel at home in the world, we do not
engage in inquiry. Consequently, no epistemological categories can fully
explain our experience, and especially not categories which assume that
experience is a commodity consisting of independent moments that directly
present themselves to us.
Suppose
that there were strong evidence from Neuroscience that human beings are not
fundamentally organisms that know? That experience was fundamentally
constituted not by an awareness (a knowing-that) of independent facts but rather by a kind of skillful coping
(a knowing-how) that could operate independently of verbalizable knowledge?
Surely such scientific evidence would strengthen and clarify Dewey's theory of
experience. Dewey believed that "experience is in truth a matter of
activities, instinctive and impulsive in their interactions with things . .
. the neglect of the deep seated
active and motor factors of experience is a fatal defect of the traditional
empirical philosophy"(Dewey 1923 p.281.) If modern neurological evidence
indicates that knowing-how and the motor factors of experience are more
fundamental than knowing-that, and could operate independently from linguistic
knowledge, wouldn't the biological structures that made such skillful coping
possible be a good candidate for the embodiment of those non-cognitive aspects
of mental life that Dewey calls experience? I think so, which is why I am now
going to lead my readers into the
thickets of connectionist cognitive science.
Connectionist Function and
Structure
I
believe that the neurally inspired structures of connectionist architecture are
the best functional descriptions we have of what Sellars called sensa or
sensations, and Dewey called Experience. These architectures are based on
neuroscientific and Artificial Intelligence (AI) research that Sellars would
probably not have had access to, and Dewey could not possibly have had access
to. But they make it possible to talk about sensations and experience as
something distinct from linguistic thought, without having to rely on mystical
concepts like immediate givenness or direct apprehension. Before I can
demonstrate this, however, I'm going to have to describe some of the ways that
a connectionist net creates organized behavior. This story has been retold many
times in the books and journals of the cognitive philosophy community. (P.M.
Churchland 1989, 1995; P.S. Churchland 1992; Clark 1993; Bechtel and Abrahamsen
1991; Bickle 1998; Haugeland 1997 and on and on.) No one who tells this story
emphasizes exactly the same points, for the full philosophical implications of
these discoveries are still being explored. My version of the story will, of
course, help to grind my particular philosophical ax, and will therefore ignore
many other points of philosophical
and scientific interest. Those who find my version too lacking in detail are
encouraged to read any and all of the other sources cited above.
I
am going to try to describe connectionist nets in purely functional terms. I will, in other words, provide a relatively
brief abstract description that makes almost no reference to the
"wet" neurology, and very little reference to the computational math,
that makes connectionist learning possible. There will consequently be very
little about physically how connectionist nets work; the emphasis will
be on what it is they do. This will hopefully enable me to avoid boring those
who have already read the neuroscience and the AI, and avoid confusing those
who have not. But more importantly,
it will bring into sharp focus those characteristics of connectionist
nets that are important for the philosophical points I am trying to make.
Specifically, I'm going to emphasize how neural nets are able to perform
certain important cognitive functions without using anything that can be called
language or logical inference. This will help show why, if knowledge/thought is
fundamentally linguistic, we can then literally give flesh to the other side of
both Dewey's distinction between knowledge and experience and Sellars' distinction between thought
and sensation.
A
connectionist net is capable of the following two behaviors which are of
interest to us.
1)
It can transform one set of numerical vectors into another set, in ways that
make possible appropriate responses to an environment. The simplest starting point for explaining connectionist vector
transformations is neither
connectionist nor a vector. Consider a thermostat, which responds to a single
numerically measurable value (the temperature of a room) and uses that
information to change some other numerically measurable value, thus controlling
a device that either increases or decreases the heat output of a furnace.
(figure 1). 
A connectionist net is able
to take not just one such input, but a large array of numerical inputs ( a
vector), and transform that vector into another vector which can in turn
perform some function in a biological system. (in much the same way that the output of the thermostat performs a
function in a heating system). The function is far more sophisticated, however,
because the connectionist net is able to take a particular group of
input vectors and respond to them with a single output vector, thus
creating a one-many relationship that functions in ways similar to the
universal-particular relationship in conceptual thinking. (figure 2) It can,
for example, learn to recognize the vector produced by light rays bouncing off
several different type faces of the letter "A" and onto an array of
light sensitive inputs. By "recognize", I mean that it will produce
the same output vector each time it encounters the many different input vectors
that are all caused by different type faces of the letter "A". In
fact, a single connectionist net can learn to make a variety of different
responses that can be used to classify a manifold of inputs into categories. It
can, for example, learn to identify many different "A"s as all being
"A"s, "B"s as
"B"s, etc. for all 26 letters of the alphabet.
2)
It can respond to new situations in ways that can be described as
'learning".
Connectionist nets are touted as "the
machines that learn", and this usually refers to two very similar virtues.
Firstly,
there is a complicated process by which one adjusts the weights between the
connections of the nodes in the net. This process is called "training the
net", because it results in a net that makes fewer errors. However, many
people claim we have no reason to assume that this training process significantly
resembles the way our own neural networks are trained, and those who claim
otherwise admit that there is still no proof for any of the best possibilities(
see Stork 1989). Consequently, although this training process is a very helpful
piece of technology, we cannot accurately extrapolate from it to our own case.
Strictly speaking, artificial connectionist nets do not teach themselves during
this process, they are shaped and taught by a very unconnectionist algorithm.
There
is, however, another quality that nets have which is very much like learning
and very much intrinsic to them. During the training process, a net is exposed
to a relatively small number of training examples, and on the basis of those it
is able to make judgments about many examples it has not yet seen. If you train
a letter recognition net to respond correctly to several examples of the letter
"A", it will also be able to correctly identify numerous other
type faces of the same letter it has never seen before, without any further
training.
Here
is a simplified explanation for this apparently amazing fact. Think of a single
neuronal node in a net as representing a dimension in a Cartesian coordinate
space, and think of the nodes' possible inputs or outputs as describing points
along that dimension. If we have ten possible values that are significant to
the network for each node, then 3 nodes could describe a thousand different
points (103) in what is called
computational space. Figure. 3 is a picture of a very simple computational space
of this sort, which is divided up into two distinct regions.
Every possible signal the net can
receive is a point in this computational space. What this particular connectionist net learned to do was
respond to any signals it received with one of two possible outputs, depending
upon which of these two regions contains the point described by value of the
signal. A net trained to recognize the letters of the alphabet would have to
make twenty six different responses, and thus the computational space would be
divided into twenty-six regions. During the training process, the net is
exposed to members of the categories it is learning to recognize, and then the
boundary lines of these computational regions are readjusted so that the net
gives the correct response in each case.
And once this training process has been completed, the net will respond correctly to other
members of the same category, because the signals they produce are within the
borders of the newly adjusted computational spaces.
Thus,
a network trained to recognize the letter "A" will see a new type
face that was not part of its original training set, but the net will recognize
it because its input vector describes a point within the boundaries marked off
during the training process. Because there could be literally billions of
points within a region that was outlined by relatively few training examples, a
properly trained network can make accurate judgments about many things it has
never seen before. Thus once our letter-recognition device has been trained to
recognize the letter "A" in Palatino, Geneva, and Zapf Chancery, it
will automatically correctly classify
"A"s in
Helvetica, Times, and Courier, because the input produced by
"A"s in those fonts
occupies the same region of computational space as the training set, as do
"A"s in hundreds of other typefaces the system has not yet seen (or
may never see, as well as possible typefaces that may never exist.) This
ability to extrapolate from a training set is greatly increased as the amount
of computational space increases. To understand just how much computational
power a biological brain
possesses, ponder instead of a
cube, a region of space that has a million dimensions. (you won't be able to
actually imagine this unless your mind is very different from mine.) This
region would have 101,000,000
points (i.e. 10 to the millionth power) that could be carved up in almost any
conceivable way[3], and each region could be associated with an
appropriate response by means of a training process. And once this association
has been established, the net would be able to make the association correctly
for billions of points it has never encountered before, because those points
are also within the borders of the demarcated space. Because the human brain
has something like 1012 neurons, it
could in principle learn to make distinctions that divide up 10 to the 1012 points, (which is much more than a trillion
trillion times as many points as there are elementary particles in the
universe) into many different spaces of many different shapes.
Given
the vast amount of resources that our nervous system has to work with, it seems
highly plausible that it could use the same basic architecture to establish a
similar relationship between A) the computational space described by the input
received from our retinas and B) an appropriate linguistic response, such as
"there is a pink ice cube in front of me". Each region of computational space could be seen as a sense
datum, and the fact that the computational space is so vast seems to explain
how we are able to store so many of them. One region of computational space
could contain the sense data, and the other could contain a list of sentences
like "there is a pink ice cube", "there is a red apple",
and so on, and learning would consist of correlating the right retinal input
with the right sentences
This
description between the relationship between language and experience might
satisfy Sellars, at least under certain interpretations. But it would not have
satisfied Dewey. He saw experience
not as discrete moments correlatable with observation sentences, but rather
divided into activities that blurred into each other, and took amounts of time
varying from seconds to hours. And more importantly, he saw experiences not as
detached observations of a world, but rather constituted by the emotions, goals and desires of
the experiencer. The Deweyan view
is a much more accurate view of experience as understood by modern
neuroscience, although sometimes certain AI experiments make it difficult to
see this. To some degree, the AI
experiments described above are shaped by the presuppositions of logical
positivism, and research has a tendency to find what it is looking for. For one thing, these experiments superimpose the law of the excluded middle
on top of a system that doesn't really support it. The outputs of the net are
not actually identical for every single response that the experimenters label
as identical. In an alphabet
recognition program, a certain range of
responses is labeled as the "A" response, another range of responses
is considered to be the "B" response and so on. When the
experimenters are training the net, they are only trying to get every single
"A" to produce a response within the "A" range. For the
purposes of the experimental protocol, these responses can be interpreted as
one of twenty six different and discrete responses, which can be correlated
with observations sentences like "That's an A", That's a B"
etc.. And this is an essential interpretation if we start with the assumption
that a knowledgable person is one who can emit the correct sentence when in the
presence of the object that the sentence refers to. The fact that no two
responses are exactly alike can be dismissed as harmless noise that doesn't
really detract from the signal.
When a neural net is in an actual
biological brain, however, the perceptual inputs it receives are rarely used
for the purpose of generating observation sentences. In fact they are never
used for that purpose in animals, for obvious reasons. Instead they are used to
control complicated pieces of skillful motor movements--tracking prey with the
eyes as it runs through tall grass, monitoring the sound of a violin to make sure that it stays in tune, or
stripping leaves off of branches in such a way as to make sure that the edible
parts come off in your prehensile paws or trunk. In order to control a process
as complicated as any of these, the neural net needs to set up a relationship
with more than just a single output vector that can be assigned a single
linguistic value. It needs to have
a relationship with another whole region of multidimensional computational
space, and that relationship needs to vary depending on the variations in the
stimulus. Thus the analog nature of the net's responses, which is deliberately
ignored when pattern recognition means making a particular linguistic response,
becomes essential to the functioning of those nets which are designed to
control "knowing-how" abilities like the examples listed above. We
would thus need to redescribe the basic architectural principle of connectionist nets to capture their
ability to embody know-how.
Instead
of correlating a cube with a set of observation sentences, we would need to
think of two cubes A and B, each like the one in figure 3, except that each
cube is divided into, say, 26
different regions instead of only 2. (figure 4)
No two regions in either cube
are the same shape or cover exactly the same territory. Now (in order to help
you visualize the computational relations between the two cubes) mentally color one region in each cube
red, another pair of regions in each cube blue, and so on for all 26 different
regions. What a connectionist net does is establish relationships between
ordered pairs of computational spaces, so that whenever the net receives input
that describes a set of points within the red region of cube A, it outputs a
set of points within the red region of cube B, and so on for all 26 colors
shared by each cube. Imagine, for example, that cube A contains
all the different points that describe a perceptual input, such as all the
different angles and speeds of an oncoming fast ball, and cube B contains all
of the different ways that one could stand, hold the bat, times ones' swing
etc. If the relationship between
the two cubes is tuned properly, the muscles' movement will change as the
perceptual inputs change, and the result will be a home run for most of the different ways that the
ball could travel over the plate. And in principle the same kind of
relationship could be established between an organism and its environment for a
variety of other skills: painting a picture, closing a business deal,
comforting a child.
Of
course skills of this level of complexity could not be modeled by only two connected
computational spaces. To get a sense of the architecture for these more
sophisticated skills, we would
have to imagine even more multidimensional spaces that are linked together in a
variety of complex ways , including links that feed back to previous links.
(Figure 5). 
There is far more to be learned about such architecture, but what
we know already gives strong evidence for a point that is important for our
purposes. In such a system, each layer of nodes could "make a
decision" about which vector within its space to output to the next layer,
based on what input it received from the previous layer. But this
"decision" would not be a logical inference, nor could anything in it
be legitimately called a word or a sentence. The nearest thing to verbal translation
of the behavior of a trained connectionist net would be the single sentence
"When you get input that resembles this, do something that
resembles that."
repeated and strung together by conjunctions. But this description would
be completely empty, and equally applicable to every connectionist net. The
"content" of the connectionist net, the information that gives it
it's cognitive power, is determined by how the computational space is divided
up, and this is expressible only in geometry, not in any language made up of
sentences.
Although
there has been some controversy over whether the regions of computational space
in a net are describable as representations, no one would claim that their interactions were
fundamentally describable as a language. Even those who believe
that a Language of Thought theory is the best description of higher cognitive
processes acknowledge that connectionism successfully implements
"lower" processes like perception and motor control. And this means
that no Language of Thought theorist denies that there are many activities that
are describable as cognitive in some sense which is clearly not linguistic. The
controversy between Language of Thought theorists and Connectionists is over
whether connectionism can eventually reduce or eliminate a Language of Thought.
No one is claiming that things could go the other way around, and we could
eventually describe Connectionist nets with a Language of Thought theory.
When
you recall that each of these "cubes" is actually an unimaginably
vast multi-dimensional computational space, and that each of these spaces can
respond to a variety of different inputs in a variety of specific ways, it
seems plausible that such a system
would be capable of very sophisticated behaviors. And it has already been shown from successful experiments in
Artificial Intelligence laboratories, that this kind of architecture can model
a variety of sophisticated "knowing-how" abilities. These abilities
have also become one of the main focuses of Cognitive Neuroscience. P.S.
Churchland's and T. Sejnowski's
"the Computational Brain" deals almost entirely with research on
modeling motor control abilities.
Traditionally motor control has not been thought of as a cognitive activity.
The assumption was that when we
moved our bodies, our brains were merely sending signals to trigger preset
patterns that ran like clockwork. (see Dreyfus 1972 pp. 235-6 for Descartes'
views on this. ) But the recent research in cognitive neuroscience research
reveals that the connectionist
systems that control our muscular skills (or even the skills possessed by
simple creatures like leeches and cockroaches) require far more flexibility and
responsiveness than is possessed by the preset algorithms that govern the
behavior of mechanical toys. This is why mechanical toys fall over much more
frequently than we do; because they do not receive feedback from the
environment. People have tried to
build robots which received feedback that was translated into some kind of
language read by a centralized processor, and then translated into motor commands. But the feedback
produced responses that were so slow as to be almost useless. (see Clark 1997)
Connectionist nets have a flexibility lacking in language based systems because
1) vector transformations can translate perceptual input directly into motor
control, without an intermediate
linguistic process 2)
Connectionist nets can give the right motor response even if the perceptual
input they are receiving is not identical to the inputs they were trained on.
The only thing they need is for the input to be appropriately similar, in a sense of similarity that is geometrically,
(not verbally), described by a region of computational space.
Knowing-How and Neural Nets
To
sum up: Connectionist nets
perform vector transformations which interrelate different regions of
computational space. This basic principle has been
effectively used to model both perception and motor control, and research in
cognitive neuroscience has revealed that it is often misleading to distinguish
the two. In most cases, what a neural net does in real biological life is use
perception to directly guide motor control. And this guidance is accomplished
by means of processes which are not at all linguistic, which means that they
make possible a kind of knowing-how that does not rely on any sort of
knowing-that. It thus seems that the following quote from Sellars states a once
widely held philosophical position that
has been pretty much disproven by what we have learned about
connectionism.
It can be argued that anything
which can be properly called "knowing how to do something"
presupposes a body of knowledge that; or to put it differently knowledge of truth or facts.
If this were so, then the statement that "ducks know how to swim" would be as metaphorical as the statement that
they know that water supports them (Sellars 1963 p.1)
We
can see from the "if" that begins the second sentence that Sellars
was clearly ambivalent about this claim himself. But it is obvious from his
wording that he believed many of his readers in 1963 would have taken it very
seriously. But now there are many people in Cognitive Science who believe that
the facts cannot possibly support such a claim. (See Dreyfus and Dreyfus 1986,
Dreyfus 1996, Clark 1993, and
Hilditch 1995). I count myself among their number, and I think that anyone who
is truly familiar with connectionist nets would agree with me. There is nothing
going on in a connectionist net that remotely resembles logical inferences from
propositions that state facts. And
yet connectionist nets are very successful at modeling both motor skills and
perceptual judgments, which are obviously the main components of the most
prototypical kinds of "knowing how". Indeed, a duck's knowing how to swim
must consist entirely of motor skills and perceptual judgments. We already know
enough about Cognitive neuroscience to be almost certain that this is true of
cockroaches (See Churchland and Sejnowski 1992, as well as the numerous
citations within it.) More
importantly, Connectionist nets are far more successful than were the attempts
at modeling perception and motor control with the logical and sentential
paradigms of symbolic systems AI. ( See Dreyfus 1994.)
Although
knowledge of the right truths or facts clearly provides aid and comfort to the developing
of "knowing-how" abilities of humans, awareness of these facts
usually doesn't come into play when the abilities are in use. Teachers of
knowing-how abilities (such as athletic coaches, and arts instructors) have
usually dealt with this fact by either 1) saying that the knowledge of the
facts drops out of consciousness, but continues to function subconsciously in a
basically linguistic manner or 2) relying on mystic exhortations along the
lines of "cultivate the mind of no-mindedness" or "use the
force!". Neither of these alternatives is necessary when we can see
linguistic abilities as distinct from perceptual and motor abilities, because
the former uses logic and sentences and the latter relies on vector
transformations. A (rather ugly) connectionist translation of "cultivate
the mind of no-mindedness" would be "don't let your linguistic mental
processes get in the way of your vector-transformation processes."
To
understand the fundamentally non-linguistic nature of vector transformations,
consider our old friend the thermostat. There is a real and meaningful sense in which a digital computer can
be said to talk to other computers, or to itself, when it manipulates bits of
information in accordance with logical principles. However, no one but a poet
(or a Disney cartoonist) would be tempted to say that a thermostat is talking
to the furnace when it translates the temperature of a room into the exact
voltage necessary to lower the fuel input into the furnace and thus eventually
restore the desired temperature. (Although some would describe this function as
representational. See Bechtel 1998) The connectionist systems that govern motor
control operate under the same basic principle as the thermostat, only they
take many different arrays of perceptual inputs and translate them into many
different arrays of motor control outputs. And because each layer in the system marks off vast regions
of computational space, each of which describes a range of possible appropriate
responses, it can respond with a level of flexibility (described by Hilditch and Merleau-Ponty as
"improvisational"[4]) which is far beyond the capabilities of a
non-connectionist translation device like the thermostat.
Both
Dewey and Sellars prophecy this distinction between the linguistic and the
non-linguistic with distinctions of their own. Dewey was far more in tune with
this modern formulation, for he saw what he called "experience" as
consisting of perception, motor control, and purposive strivings united in a
single organic whole. Sellars, as we shall see in the next few paragraphs,
originally saw his concept of "sensation" as a component only of
perception. To some degree, he started out assuming that he would only need to
make a few reformations of classical sense datum theory, but his reforms
eventually became every bit as revolutionary as Dewey's.
Sellars and Connectionism
In
"the Structure of Knowledge", Sellars starts with the sense datum
theorist's paradigm of knowledge: The single experience of deliberately looking
at an object in front of us. He then claims this apparently unified experience
consists of two distinct process, and (as he often said in class) the only
reason that these processes occur simultaneously is that people can chew gum
and walk at the same time. The first process is completely cognitive, and
consists of the occurrence of a sentence in the mind, saying something like
"lo, a pink ice cube over there". The second process is completely
non-cognitive, and consists of a manner of sensing. At the time one would be
thinking the previously mentioned sentence, there would also be an occurrence
in the mind of a process Sellars
called "sensing a-pink-cube-ly". From what we now know about
connectionism, it is clear that Sellars was largely right about this
distinction, because the processes of vector transformations that create
sensory experience are fundamentally non-linguistic, and for philosophers of
Sellars' generation all cognition was linguistic. Even though Sellars
frequently expressed misgivings about this assumption, he never really felt
that he had the resources to offer any alternative.
In
spite of this lack of resources, however, he comes extremely close to making a
distinction between two distinct kinds of knowledge which is coextensive with
the language/vector transformation distinction, even though it could not
possibly contain any of the computational details. In "the Structure of
Knowledge" Sellars introduces his distinction between thought and sensing
by saying that musicians and composers have two different ways of thinking
about their art. They can think about sound (i.e. linguistically) and
they can also non-linguistically think in sound. He then makes the
following conclusion from this, which seems to contradict many of his other
statements.
There
is much food for thought in these reflections. . . But the fundamental problems
which they pose arise already at the perceptual level. For as we shall see,
visual perception, itself is not just a conceptualizing of colored objects
within the visual range-a 'thinking about' colored objects in a certain
context--but in a sense most difficult to analyze, a thinking in color about colored objects (Chastened Ed. 1975 p.305)
Sellars
was certainly right that this sense was most difficult to analyze. How can the
above be reconciled with this claim?
all awareness of sorts,
resemblances, facts, etc., in short all awareness of abstract entities --
indeed, all awareness of even particulars -- is a linguistic affair (Sellars 1963).
Is there anything left to be aware of
that isn't on the above quoted list? And what do we do with his repeated claims
that sensations are non-cognitive? The first paragraph of Sellars 1981 actually
seem to be implying that sensations are also different from non-linguistic
thought. This seems to leave us with three categories of mental events,
exemplified by 1) linguistic thoughts about sound 2) non-linguistic thinkings in
sound, and 3) Audial sensations. There is no question that Sellars worked very
hard at this question, and I won't claim that there isn't an argument buried in
his writings somewhere that exonerates him of logical inconsistency on this
point. But now that we have the conceptual resources of connectionism to explain
what happens during "Thinking-in-sound" or
"Thinking-in-shape-and-color", I think it makes sense to cut this
particular Gordian knot rather than continue to try to untie it. The most
plausible exegesis is that Sellars was right when he said that sensing is
non-linguistic, and wrong when he said it is non-cognitive. This is perfectly
understandable mistake, given that linguistic cognition was the only kind he
knew about. For Sellars, (and his opponents) there were only two alternatives:
Awareness was either linguistic or direct apprehension of something immediately
given, and he knew better than anyone that the latter alternative was
unacceptable.
There
is also something very right about Sellar's insistence that the sensing part of
the perception was non-inferential, and it became the basis of one his most
important critiques of sense datum theory. It really is impossible to make an
inference from a particular "sensing a-pink-cube-ly", even though it
is fully justifiable to make an inference from the mental sentence that accompanies it. To think otherwise (as
did the sense datum theorists) is to make a category mistake along the lines of
trying to make an inference from a walnut. (as opposed to making inferences
from a sentence about a walnut.) This is why Sellars describes our sensations
as being "non-inferential." However, the fact that you cannot make
logical inferences from a sensation does not mean that it is not cognitive in
the broad sense that connectionist cognitive science defines "cognitive".
For if a sensation is the product of a series of appropriate vector
transformations it might enable us to move skillfully through the world even if
it was not accompanied by any sentences that would enable us to talk to
ourselves about it. It might, for example, help me in reaching skillfully
across the room to pick up the pink ice cube and put it in my drink, even
though I was completely absorbed in the discussion of some other topic.
Sense Data vs. Sensory Process
If sensing and thinking can be divided
this way--if they are as distinct as chewing gum and walking-- wouldn't there
be many occasions when we would chew gum and not walk, or vice versa? Although Sellars acknowledges this in
principle he almost always used the same paradigm explanandum as the sense
datum theorists: a single moment of perception, in which one focuses on an
object in one's perceptual field and makes a judgment about it. He did not
slice perceptions as finely as did the classical sense datum theorists, who saw
the fundamental parts of perceptions to be rather like frames of film that
contained weird artifacts like elliptical pennies. But the examples he usually
used required him to share the sense datum theorist's presumption that the world is always encountered as what
Heidegger called being-present-at-hand. I sit in my chair, the pink ice cube
sits on my table, and I look at the pink ice cube in an attempt to understand
it, or acquire some information about it. The sentence that comes to me at the
moment of perceptual taking lasts a second or two, and there is a sensation of
color-and-shape going on during the whole time I am thinking that sentence.
From this, it seems tempting to infer that Sellars sees our sensations as
divided up into chunks of a second
or two, each of which matches one
of those mentalese speech acts that he calls a perceptual taking. Consider, for
example, this passsage from Sellars 1963, which offers an explanation for how
linguistic knowledge can be derived from pre-linguistic experience.
While Jones's ability to give
inductive reasons today is built on a long
history of acquiring and manifesting verbal habits in perceptual situations,
and, in particular, the occurrence of verbal episodes, e.g. "This is
green," which is superficially like those which are later properly said to
express observational knowledge, it does not require that any episode in this
prior time {i.e. before Jones had language} be characterizeable as expressing
knowledge.
. . . {Footnote added in 1963} My
thought was that one can have direct (non-inferential) knowledge of a past fact
which one did not or even (as in the case envisaged) could not conceptualize at
the time it was present. (Sellars
1963 p.169)[5]
I
first interpreted this passage to mean that if a baby looked into a refrigerator
before she ever could say to herself "that is a Pink Ice Cube", the
experience produced in her would still be divided up into cubes of pink,
tray-shaped chunks of tray-color, and "Mrs. Paul's frozen fish
sticks"- shaped chunks of "Mrs. Paul's frozen fish
sticks"-color. If my interpretation had been correct, Sellars would have
been a reformer of sense datum theory, rather than a revolutionary working
against it. To paraphrase his words for my purposes, he would be teetering on
the brink of sense datum theory, and claiming that he has not fallen over the
edge. And as one can no doubt tell from the sarcastic tone of my examples, I
would not have been sympathetic to this view.
But
many years later, with the appearance of his three Carus lectures, he ends up
not only stepping back from the brink, but running so far away from it that
almost no one in the analytic tradition is willing to follow him. And as Jay
Rosenberg 1982 points out, there are many indications (including the quote from
Sellars 1963b that closes the Carus lectures) that Sellars believed something
like these revolutionary claims even when he described himself as basically a
reformer of sense datum theory. To sum up the Carus lectures in the most
flamboyant way possible: Our sensory experience is a pure absolute process. It
ebbs and flows as "an on-going tissue of goings-on" with moments that
can be conveniently given labels like "sensing a-pink-cube-ly".
Fundamentally, however, there are no objects at all. Even the objects of
science should seen as examples of the pure process "quarking" and
"electroning". This, Sellars claims is the only way we can account
for the fact that there are occurrent properties like shape and color sharing
the universe with theoretical entities like quarks and electrons.
This
view resembles both Dewey and James in many interesting ways, and Sellars'
arguments for it make it far more convincing than it appears to be in this
blunt statement. But from this God's eye view of the Ultimate Nature of Things,
it's rather difficult to consider the issues of epistemology and philosophy of
mind that are the main focus of this paper. And more importantly, the
unwillingness to accept anything less than such a God's eye view is the crucial
difference between Sellars' Scientific Realism and Dewey's Pragmatic Pluralism.
Before Sellars puts forth his theory of absolute process, he describes four
other alternatives, all of which he claims are unacceptable. Two of them, one
of which he calls instrumentalism , and the other emergentism, pretty much cover Dewey's metaphysical position. The following
passage from EPM nicely sums up Sellars' differences and similarities with
Dewey on this issue.
As long as the existing framework
{i.e. of the common-sense Manifest Image} is used, it will be incorrect to say -- otherwise than to make a philosophical point about
the framework --
that no object is really colored, or is located in Space, or endures through
Time. But, speaking as a philosopher, I am quite prepared to say that the common-sense
world of physical objects in Space and Time is unreal -- that is, that there
are no such things. Or, to put it less paradoxically, that in the dimension of
describing and explaining the world, science is the measure of all things, of
what is that it is, and of what is not that it is not. (Sellars 1963a p. 173
italics in original)
There
are certain hints of something like pragmatic pluralism in the first sentence
quoted above[6], as there are in Sellars' definition of success
in philosophy as knowing ones way around with respect to all things (Sellars
1963b p.1) But the rest of the
quote makes it clear that for Sellars, science is strictly speaking the measure
of all things. Dewey rejected the
idea that any one system of thought could be the measure of all things.
He believed that there were a variety of ways to measure and comprehend
things, and the question of which
was the best was answerable only if you qualified it by answering the further
question "best for what?".
Consequently I don't think Dewey would claim that Sellars is mistaken in
claiming that the universe is an
absolute process, only in claiming that this absolute process was realer than
the various relative processes that make up the daily flux of our experience.
If your goal is to resolve certain philosophical problems, Sellars' absolute
process theory may be arguably best for that purpose. But according to Dewey,
that does not make it the measure of all things, it just makes it a very good
theory for philosophical purposes. A scientific realist like Sellars would
probably consider such a view to be instrumentalist, although it does not
entail many of the instrumentalist assumptions that are associated with instrumentalists in the analytic
tradition like Bas Van Fraassen and the operationalist Carnap. (Sellars
outlines some of these assumptions on pp. 43 and 44 of his 1963b.) And because
Dewey believes that there is no one answer to the question "what is
truth?", the whole problem of emergence doesn't even come up. Emergence is
only a mystery if you think that the universe must be constituted by a single
unified system. If so, when something emerges that has no obvious relationship
to that system, this is a fact that calls for an explanation. But there is no
puzzle of figuring out how the world ties together into a single system if the
world is not in fact tied together into a single system. [7]
Consequently,
for Dewey the key issue is not the
relationship between some
kind of pure process and knowledge. It is the relationship between experience
(which is constituted by processes that are directed by our goals and emotions)
and inquiry (which is a way of transforming those process in such a way as to
enable us to better know our way around the world in which we live.) And as
with Sellars, I think that a consideration of the connectionist nature of our
minds will clarify much of what previously appeared to be mysterious in Dewey's
theory of experience.
Connectionism
and Dewey's
concept of Experience
In his 1994 article, Shusterman criticizes Dewey's concept of experience as a kind
of unacknowledged foundationalism, because he claims Dewey sees immediate
quality as "the underlying guide of all thought and discourse"
(p.134). Shusterman claims that Dewey asserts this about all thinking,
discourse and judgment, and that this amounts to "a foundational
metaphysics of presence justified by transcendental argument" (ibid.) He
suggests that Dewey can remain truer to his own spirit if he acknowledges that "given the structuring of habit
and purpose, there is no need to appeal to a special unifying quality, a felt
presence, to organize the situations as an organized whole."(ibid.) Shusterman claims that anything that
can be clarified by discussing qualitative thought or experience can be dealt
with more effectively by considering how our behavior is conditioned by our
habits and purposes, without leaving us open to the dangers of foundationalism
and/or ineffablity. He does not feel that Deweyan philosophers should ignore
immediate experience. But he feels that it would be more in line with Dewey's
principles to use philosophy to "establish and improve the quality of
immediate experience as a practical and useful tool." (p.136), rather than
using it as a theoretical foundation in epistemology.
My
debt to Shusterman in this paper is enormous, and it is one of the pathologies
of the philosophical profession that I can only acknowledge this debt by
criticizing his article. The
reader should remember that I am taking the trouble to show where I think he is
wrong because no one else has gotten so much else right. I could quarrel with
his conflation of benign concepts like "guide" with pernicious ones
like "foundation". But I think most of my disagreements with him
spring from the fact that he assumes no one in the contemporary philosophical
community would question Pierce's claim that "my language is the sum total
of myself". To some degree he is right, but this consensus rests on the
assumption that Quine and Sellars have proven this point, and as I tried to
show earlier in this paper, this is not all that obvious. Shusterman quotes
Sellars as saying "All awareness is a linguistic affair" but makes no
reference to any of Sellars' equivocations on the issue of sensations and
"thinkings-in-color". In fact, while Dewey may be out of step with
Sellarsians like Rorty, his position on experience is very similar to that of
Sellars himself.
. . . .to be a smell is one
thing, to be known as a smell, another; to be a "feeling" one thing,
to be known as a "feeling" another (Dewey 1910 p.81)
If
we substitute
"a-pink-cube"ly manners of sensing for smells and feelings,
this claim is identical with the Sellarsian distinction between thoughts and
sensations we discussed earlier. This means that connectionist theory should be
as helpful in clarifying Dewey's theory of experience as it was for Sellars'
theory of sensation. I maintain
that talk about sensations and/or qualitative experience appears to refer to
some kind of mystical ineffability only when the sole model we have for mental
life is a linguistic one. If the
smells and the "feelings"
are constituted by vector transformations of the sort described previously,
and the thoughts (which make knowledge possible) are constituted by inner
linguistic events, the ineffability is no longer mysterious. What we have
instead are two completely biological processes taking place in living
organisms. But because one of the processes is linguistic and the other is not,
the latter cannot be reduced to the former. Dewey's unique contribution was to
show the advantages of assuming that they were different processes.
Unfortunately, because he could not explain the two different mechanisms that
distinguished knowing from feeling, it sounded like he was being mystical when
he claimed they must be separate.
Until
very recently, it was usually assumed that both sensations and thoughts,
knowings-how and knowings-that, were constituted by processes that were
essentially the same and essentially linguistic and propositional. (As Sellars
was fond of pointing out, the classical empiricist mistake was to assume that
sense data granted an awareness of facts, from which logical inferences could
be made.) Jerry Fodor ridiculed this idea by saying that when we tie our shoes,
there is a little man in our heads who gets a book down from a shelf titled
"how to tie your shoes" and then follows the instructions in the
book. (Fodor 1981) But as is his wont, after ridiculing this idea, Fodor then
claims that it is the only real explanation we have, and that it is
fundamentally sound. According to this theory, the difference between those
processes that we use in skillful coping and those that we use in verbal
deliberations is nothing more than the fact that the latter are under the
spotlight of awareness and the former are not.
Dewey
wanted to claim that experience is not just vaguely perceived knowledge, but
something different in kind from knowledge; something constituted by our
habits, skills, and abilities, and necessarily linked to our goals,
aspirations, and emotions. Throughout it's history, philosophy has concentrated
on understanding the architecture that made rational thought possible: i.e.
language. Language is the thing which separate us from the brutes, and entitles
us to the Aristotelian honorific "Rational Animal". For that reason,
most philosophers associated conceptual thought with the abstract and the
spiritual, and it was thus considered to be the thing that made us conscious
beings. Felt experiences were considered to be thoughts that were deficient in
some way: They were confused thoughts, waiting to be brought into focus, or
atomistic bits of thought waiting to be assembled into scientific theory.
Dewey's radical claim was that experience was not only distinct from thought, but
also had rules of organization that were all it's own.
I've
devoted quite a bit of space in this paper to explaining some of the
capabilities of connectionist nets, in hopes of demonstrating how they can
justify two closely related distinctions in Dewey and Sellars. There are also
some other developments in experimental psychology which provide even more
empirical justification for Dewey, and relate significantly to connectionism as
well. When experimental psychology began to encounter the inadequacies of
traditional sense datum theory, psychologists like J.J. Gibson and Ulrich
Nesser began to talk about considering the senses to be organized into
perceptual systems. (Gibson 1950, 1966. Nesser 1992, 1994). Gibson, in
particular, made some very perceptive observations about certain organizing
principles that govern the perceptual field, which were a welcome relief from
the traditional view that sensory experience consisted of independent directly
given particles. Gibson's The
Senses Considered as Perceptual Systems seems to imply that these
perceptual systems present themselves to consciousness in essentially the same
way that the sense data were once assumed to do. It appeared that the only
difference was that the perceptual systems presented themselves all at once
rather than in little bits, or perhaps that they provided a form that
conditioned the raw content of sensations. However, there is no such implication in Gibson's later
work, such as the Ecological approach to Visual Perception, which
describes a space not filled with neutral objects, but divided up into what he
calls affordances. In an essay
of his quoted in Hilditch 1995 (p.42), Gibson describes the space we experience
while driving an automobile. What we experience is not shapes and colors, or
even objects, but specific opportunities for driving action: the field of safe
travel, the minimum stopping zone, the distance required to reach from the
brake to the clutch. Each of these fields is called an affordance, because it
affords an opportunity for a specific kind of action.
Although
he uses the term "direct perception" to describe the relationship
between organism and affordance, Gibson means almost exactly the opposite of
what the sense datum tradition meant by that term. For Gibson, the perception
of an affordance is direct because it does not make a stop in the cognitive
center of the brain long enough for us to construct a sentence about it.
Instead it goes directly from the retina into muscular control centers,
bypassing the kind of conceptual categorization that makes thought and
reasoning possible. Although Gibson did most of his work before connectionist
theory was developed, he did a thorough job of observing the sort of functions
that connectionism is good at implementing. The network of affordances that
defines our relationship to our world is
constituted by opportunities to utilize "knowing-how" skills
and abilities. And Gibson was able
to see that such a network functioned without reference to linguistic thought,
even before connectionist theory existed to explain exactly how that was
possible.
Gibson's
claim that our primary experience is a world of affordances is extremely
similar to Dewey's. Gibson's claim that our awareness of affordances is direct,
and distinct from the kind of knowledge that makes categorization possible,
seems to imply that we can map his distinction between direct and categorizing
perception onto Dewey's distinction between knowledge and experience.[8] The only apparent problem with this mapping is
that Dewey considered experience to contain not just skills and abilities, but
also goals and emotions. I think, however, that extending Gibson's theory in
this direction is a very natural extrapolation for many reasons. To see
something as an affordance is to see it as potentially useful for some purpose
or goal. Dewey believed that it was our goals that caused our experience to be
tinged with a distinct emotional quality, and saw that quality as being the
starting point of all inquiry. If we recall that no one ever drives a car
unless she plans to go somewhere, and no one goes somewhere unless she has a
reason, We see there is no reason to stop the structuring principle of
affordances at the level of simple motions like walking and balancing. The act
of driving is itself an affordance to getting somewhere, and so on. As a
scientist, Gibson was obligated to confine himself to the study of those
relatively primitive affordance structures that could be studied in laboratory
conditions. But if we are to confront the traditional problems of epistemology
with new solutions, we have to be willing to extrapolate to the higher
cognitive and/or experiential functions that are the subject matter of
epistemology. And when we do this with the concept of affordances, we end up with
something very much like Dewey's concept of experience.
In
real life skills are never separated from goals, and goals are never separated
from emotions. It is, of course, necessary to separate them for purposes of
analysis, for a given skill can be practiced in service of a variety of
different goals, and we can feel many different emotions about the same goal.
We can hammer in a nail to build a house or build a coffin, and we can feel
either happy or sad about building either or both. But no one ever hammers in a
nail for no reason at all (even hammering in a nail to prove that it was
possible to hammer in a nail for no reason, would be doing it for that reason.)
And whatever emotions we feel about anything are conditioned by how that
occurrence affects our goals. If we have conflicting emotions about an
occurrence, it is because we have conflicting goals. As far as I know, we don't
have enough knowledge about the nervous system, or enough success with AI
modeling projects, to provide any real evidence that emotions and goals are
constituted by connectionist vector transformations. But given that they seem
to flow so seamlessly into skills and abilities, it seems likely that they both share a common biological
architecture. And if they did, it
would explain why emotions, and the skilled habitual responses that are shaped
by our goals and emotions, seem to have an independent existence from our
verbalizable thoughts.
For
those who take ordinary language seriously, I would point out that the word
"feelings" refers to both sensations and emotions, and that we feel
far more pain at the death of a loved one than at the pricking of our fingers.
What Dewey is suggesting is that this original seamless concept of
"feeling" is the true one, and that the separation of "physical feelings" that
supposedly tell us about the world from "emotional feelings" that
supposedly tell us only about our mental states is a philosopher's artifact.
Gibson gives strong evidence that a particular retinal or proprioceptive signal
does not create a feeling all by itself. It only becomes part of our mental
life when it becomes part of a system of affordances. The only real difference
between Dewey and Gibson on this point is that Dewey is willing to see the most
sophisticated and uniquely human aspects of experience as also being systems of
affordances: Not only driving a car, as Gibson suggested, but listening to or
performing in a symphony; cooking or eating a fine meal in a restaurant. And
because Dewey did not have to fear professional ostracism if he talked about
those aspects of human experience that are not describable as behavior, he was
able to account for nuances that Gibson was professionally compelled to avoid.
This is why he was able to talk about a single goal directed activity having a
unifying felt quality, whose existence we can acknowledge (but not communicate
or capture) with single ejaculative expression like "good!" or
"alas". (Dewey 1931 p.101). And when the goal is abandoned or
achieved, the experience comes to an end, and another one with a different
quality seamlessly begins.[9]
Once
we have abandoned the hope that we can get experience to act as a foundation
for knowledge if we divide it up into the smallest possible bits, we can see
what should have been obvious to us all along: Each of the previously listed
experiences presents itself to us as a whole. Each of these wholes is divisible,
of course, but that does not mean they were divided when we first
experienced them. That this is
true even of experiences which are of relatively short duration can be seen by
the following example, familiar to any musician. If one learns a four bar
melody, and there is something wrong with one's execution of one part of it, it
is usually necessary to divide it into smaller parts, so as to inquire into
exactly what is wrong. If we learned melodies by storing the notes one at a
time and then stringing them together, it should be easy to divide a melody in
half with a simple internal command. In fact, playing half the melody is like
learning a whole new melody, and when we finally do learn it, it is because we
have grasped the noticeably
different qualitative feel of the newer shorter melody. Similarly, most people
who can drive cars could not rattle off the list of affordances that the car
driving experience reveals to someone like Gibson who inquires into the
experience. And the fact that car driving experience really is constituted by
those affordances doesn't mean we were experiencing it that way all along. It
just means that the experience of driving the car was the starting point for
Gibson's inquiry, it does not mean that it provides a foundation for
that inquiry. The experience does
not remain firm and fast, the way a sense datum supposedly did. Inquiring into
the experience will change it, hopefully for the better; that is why we conduct
inquiries.
I
have said relatively little about inquiry, even though it is the other half of
the dichotomy that is the central theme of this paper. And what I have said may
have given the misleading impression that there is an unbridgeable difference
in kind between inquiry and experience. In fact, knowing how to conduct an
inquiry is a skill that must be fundamentally governed by similar principles
that govern other knowing-how abilities. And conducting an inquiry is itself an
experience that has a quality all its own, a quality often very difficult to
fully appreciate for those of us who live so close to it. It seems inevitable that eventually someone
will figure out how to apply the
principles of Gibson and the Connectionists to show how a system of affordances
can be built up that makes inquiry possible, and a few steps have already been
made in that direction.[10] The focus for such a meta-inquiry would certainly
be that ability which makes us so much better at inquiry than any other animal
i.e. language. This search is the philosopher's stone (and hopefully,
eventually, the Rosetta stone) of connectionist research, and if successful, it
would blur the distinction I have tried so hard to clarify in this paper. But
such a blurring would be appropriate for that sort of inquiry, even if it would
confuse the presuppositions of this one. For most of philosophy's history, all
human experience has been seen as a kind of continuous spontaneous knowing, so
perhaps it is high time that philosophers like Dewey, Heidegger and
Merleau-Ponty have turned the tables, and now see knowing as a particular kind
of experienced activity, rather than the only kind of experience there is.