Patricia Smith Churchland - Ricerca sull'intelligenza artificiale specializzata al MIT

Patricia Smith Churchland

How do neurons know?

My knowing anything depends on my
neurons–the cells of my brain.1 More
precisely, what I know depends on the
speci½c con½guration of connections
among my trillion neurons, on the neu-
rochemical interactions between con-
nected neurons, and on the response
portfolio of different neuron types. Tutto
this is what makes me me.

The range of things I know is as di-
verse as the range of stuff at a yard sale.
Some is knowledge how, some knowl-
edge that, some a bit of both, and some
not exactly either. Some is fleeting, some
enduring. Some I can articulate, ad esempio
the instructions for changing a tire,
some, such as how I construct a logical
argument, I cannot.

Some learning is conscious, some not.

To learn some things, such as how to

Patricia Smith Churchland is ucPresident’s Pro-
fessor of Philosophy and chair of the philosophy
department at the University of California, San
Diego, and adjunct professor at the Salk Institute.
She is past president of the American Philosophi-
cal Association and the Society for Philosophy
and Psychology. Her latest books are “Brain-
Wise: Studies in Neurophilosophy” (2002) E
“On the Contrary: Critical Essays, 1987–1997”
(with Paul Churchland, 1998).

ride a bicycle, I have to try over and
Sopra; by contrast, learning to avoid eat-
ing oysters if they made me vomit the
last time just happens. Knowing how to
change a tire depends on cultural arti-
facts, but knowing how to clap does not.
And neurons are at the bottom of it all.

How did it come to pass that we know
anything?

Early in the history of living things,
evolution stumbled upon the advantages
accruing to animals whose nervous sys-
tems could make predictions based upon
past correlations. Unlike plants, who
have to take what comes, animals are
movers, and having a brain that can
learn confers a competitive advantage in
½nding food, mates, and shelter and in
avoiding dangers. Nervous systems earn
their keep in the service of prediction,
E, to that end, map the me-relevant
parts of the world–its spatial relations,
social relations, dangers, and so on. E,
Ovviamente, brains map their worlds in
varying degrees of complexity, and rela-
tive to the needs, equipment, and life-
style of the organisms they inhabit.2

1 Portions of this paper are drawn from my
book Brain-Wise: Studies in Neurophilosophy
(Cambridge, Massa.: con la stampa, 2002).

2 See Patricia Smith Churchland and Paul M.
Churchland, “Neural Worlds and Real Worlds,"
Nature Reviews Neuroscience 3 (11) (novembre
2002): 903–907.

Dedalo Inverno 2004

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
D
UN
e
D
UN
R
T
io
C
e
–
P
D

F
/

1
3
3
1
4
2
1
8
2
8
7
4
4
0
0
1
1
5
2
6
0
4
7
7
2
7
4
6
6
8
4
P
D

B
sì
G
tu
e
S
T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

How do
neurons
know?

Thus humans, dogs, and frogs will repre-
sent the same pond quite differently. IL
human, Per esempio, may be interested
in the pond’s water source, the potability
of the water, or the potential for irriga-
zione. The dog may be interested in a cool
swim and a good drink, and the frog, in un
good place to lay eggs, ½nd flies, bask in
the sun, or hide.

Boiled down to essentials, the main
problems for the neuroscience of knowl-
edge are these: How do structural ar-
rangements in neural tissue embody
knowledge (the problem of representa-
zioni)? How, as a result of the animal’s
experience, do neurons undergo changes
in their structural features such that
these changes constitute knowing some-
thing new (the problem of learning)?
How is the genome organized so that the
nervous system it builds is able to learn
what it needs to learn?

The spectacular progress, during the
last three or four decades, in genetics,
psicologia, neuroethology, neuroem-
bryology, and neurobiology has given
the problems of how brains represent
and learn and get built an entirely new
look. In the process, many revered para-
digms have taken a pounding. From the
ashes of the old verities is arising a very
different framework for thinking about
ourselves and how our brains make
sense of the world.

Historically, philosophers have debat-

ed how much of what we know is based
on instinct, and how much on experi-
ence. At one extreme, the rationalists ar-
gued that essentially all knowledge was
innate. At the other, radical empiricists,
impressed by infant modi½ability and
by the impact of culture, argued that all
knowledge was acquired.

Knowledge displayed at birth is obvi-
ously likely to be innate. A normal neo-
nate rat scrambles to the warmest place,

latches its mouth onto a nipple, and be-
gins to suck. A kitten thrown into the air
rights itself and lands on its feet. A hu-
man neonate will imitate a facial expres-
sion, such as an outstuck tongue. Ma
other knowledge, such as how to weave
or make ½re, is obviously learned post-
natally.

Such contrasts have seemed to imply
that everything we know is either caused
by genes or caused by experience, Dove
these categories are construed as exclu-
sive and exhaustive. But recent discover-
ies in molecular biology, neuroembryol-
ogy, and neurobiology have demolished
this sharp distinction between nature
and nurture. One such discovery is that
normal development, right from the ear-
liest stages, relies on both genes and epi-
genetic conditions. Per esempio, a fe-
male (xx) fetus developing in a uterine
environment that is unusually high in
androgens may be born with male-look-
ing genitalia and may have a masculin-
ized area in the hypothalamus, a sexually
dimorphic brain region. In mice, IL
gender of adjacent siblings on the pla-
cental fetus line in the uterus will affect
such things as the male/female ratio of a
given mouse’s subsequent offspring, E
even the longevity of those offspring.
D'altra parte, paradigmatic in-
stances of long-term learning, ad esempio
memorizing a route through a forest, Rif-
ly on genes to produce changes in cells
that embody that learning. If you experi-
ence a new kind of sensorimotor event
during the day–say, Per esempio, you
learn to cast a ½shing line–and your
brain rehearses that event during your
deep sleep cycle, then the gene zif-268
will be up-regulated. Improvement in
casting the next day will depend on the
resulting gene products and their role in
neuronal function.

Infatti, ½ve important and related dis-
coveries have made it increasingly clear

Dedalo Inverno 2004

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
D
UN
e
D
UN
R
T
io
C
e
–
P
D

F
/

1
3
3
1
4
2
1
8
2
8
7
4
4
0
0
1
1
5
2
6
0
4
7
7
2
7
4
6
6
8
4
P
D

B
sì
G
tu
e
S
T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

Patricia
Smith
Churchland
SU
apprendimento

just how interrelated ‘nature’ and ‘nur-
ture’ are, E, consequently, how inade-
quate the old distinction is.3

Primo, what genes do is code for pro-
teins. Strictly speaking, there is no gene
for a sucking reflex, let alone for female
coyness or Scottish thriftiness or cogni-
zance of the concept of zero. A gene is
simply a sequence of base pairs contain-
ing the information that allows rna to
string together a sequence of amino
acids to constitute a protein. (This gene
is said to be ‘expressed’ when it is tran-
scribed into rna products, some of
Quale, in turn, are translated into pro-
teins.)

Secondo, natural selection cannot di-
rectly select particular wiring to support
a particular domain of knowledge. Blind
luck aside, what determines whether the
animal survives is its behavior; its equip-
ment, neural and otherwise, underpins
that behavior. Representational prowess
in a nervous system can be selected for,
albeit indirectly, only if the representa-
tional package informing the behavior
was what gave the animal the competi-
tive edge. Hence representational so-
phistication and its wiring infrastructure
can be selected for only via the behavior
they upgrade.

Third, there is a truly stunning degree
of conservation in structures and devel-
opmental organization across all verte-
brate animals, and a very high degree of
conservation in basic cellular functions
across phyla, from worms to spiders to
humans. All nervous systems use essen-
tially the same neurochemicals, E
their neurons work in essentially the
same way, the variations being vastly
outweighed by the similarities. Humans

3 In this discussion, I am greatly indebted to
Barbara Finlay, Richard Darlington, and Nich-
olas Nicastro, “Developmental Structure in
Brain Evolution,” Behavioral and Brain Sciences
24 (2) (April 2001): 263–278.

have only about thirty thousand genes,
and we differ from mice in only about
three hundred of those;4 meanwhile,
we share about 99.7 percent of our genes
with chimpanzees. Our brains and those
of other primates have the same organi-
zation, the same gross structures in
roughly the same proportions, the same
neuron types, E, so far as we know,
much the same developmental schedule
and patterns of connectivity.

Fourth, given the high degree of con-
servation, whence the diversity of multi-
cellular organisms? Molecular biologists
have discovered that some genes regu-
late the expression of other genes, E
are themselves regulated by yet other
genes, in an intricate, interactive, E
systematic organization. But genes (via
rna) make proteins, so the expression
of one gene by another may be affected
via sensitivity to protein products. Addi-
tionally, proteins, both within cells and
in the extracellular space, may interact
with each other to yield further contin-
gencies that can ½gure in an unfolding
regulatory cascade. Small differences in
regulatory genes can have large and far-
reaching effects, owing to the intricate
hierarchy of regulatory linkages between
them. The emergence of complex, inter-
active cause-effect pro½les for gene ex-
pression begets very fancy regulatory
cascades that can beget very fancy or-
ganisms–us, Per esempio.

Fifth, various aspects of the develop-
ment of an organism from fertilized egg
to up-and-running critter depend on
where and when cells are born. Neurons
originate from the daughter cells of the
last division of pre-neuron cells. Wheth-
er such a daughter cell becomes a glial
(supporting) cell or a neuron, and which
type of some hundred types of neurons

4 See John Gerhart and Marc Kirschner, Cells,
Embryos, and Evolution (Oxford: Blackwell,
1997).

Dedalo Inverno 2004

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
D
UN
e
D
UN
R
T
io
C
e
–
P
D

F
/

1
3
3
1
4
2
1
8
2
8
7
4
4
0
0
1
1
5
2
6
0
4
7
7
2
7
4
6
6
8
4
P
D

B
sì
G
tu
e
S
T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

How do
neurons
know?

the cell becomes, depends on its epige-
netic circumstances. Inoltre, IL
manner in which neurons from one area,
such as the thalamus, connect to cells in
the cortex depends very much on epige-
netic circumstances, per esempio., on the sponta-
neous activity, and later, the experience-
driven activity, of the thalamic and corti-
cal neurons. This is not to say that there
are no causally signi½cant differences
between, for instance, the neonatal suck-
ing reflex and knowing how to make a
½re. Differences, obviously, there are.
The essential point is that the differ-
ences do not sort themselves into the
archaic ‘nature’ versus ‘nurture’ bins.
Genes and extragenetic factors collabo-
rate in a complex interdependency.5
Recent discoveries in neuropsychology

point in this same direction. Hitherto, Esso
was assumed that brain centers–mod-
ules dedicated to a speci½c task–were
wired up at birth. The idea was that we
were able to see because dedicated ‘visu-
al modules’ in the cortex were wired for
vision; we could feel because dedicated
modules in the cortex were wired for
touch, and so on.

The truth turns out to be much more

puzzling.

Per esempio, the visual cortex of a
blind subject is recruited during the
reading of braille, a distinctly nonvisual,
tactile skill–whether the subject has ac-
quired or congenital blindness. It turns
fuori, Inoltre, that stimulating the
subject’s visual cortex with a magnet-
induced current will temporarily impede
his braille performance. Even more re-
markably, activity in the visual cortex
occurs even in normal seeing subjects
who are blindfolded for a few days while

5 See also Steven Quartz and Terrence J. Sej-
nowski, Liars, Lovers, and Heroes (New York:
William Morrow, 2002).

learning to read braille.6 So long as the
blindfold remains ½rmly in place to pre-
vent any light from falling on the retina,
performance of braille reading steadily
improves. The blindfold is essential, for
normal visual stimuli that activate the
visual cortex in the normal way impede
acquisition of the tactile skill. For exam-
ple, if after ½ve days the blindfold is re-
moved, even briefly while the subject
watches a television program before
going to sleep, his braille performance
under blindfold the next day falls from
its previous level. If the visual cortex can
be recruited in the processing of nonvi-
sual signals, what sense can we make of
the notion of the dedicated vision mod-
ule, and of the dedicated-modules hy-
pothesis more generally?

What is clear is that the nature versus
nurture dichotomy is more of a liability
than an asset in framing the inquiry into
the origin of plasticity in human brains.
Its inadequacy is rather like the inade-
quacy of ‘good versus evil’ as a frame-
work for understanding the complexity
of political life in human societies. È
not that there is nothing to it. But it is
like using a grub hoe to remove a splin-
ter.

An appealing idea is that if you learn

something, such as how to tie a trucker’s
knot, then that information will be
stored in one particular location in the
brain, along with related knowledge–
Dire, between reef knots and half-hitches.
Questo è, after all, a good method for stor-
ing tools and paper ½les–in a particular
drawer at a particular location. But this
is not the brain’s way, as Karl Lashley
½rst demonstrated in the 1920s.

6 See Alvaro Pascual-Leone et al., “Study and
Modulation of Human Cortical Excitability
with Transcranial Magnetic Stimulation,” Jour-
nal of Clinical Neurophysiology 15 (1998): 333–
343.

Dedalo Inverno 2004

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
D
UN
e
D
UN
R
T
io
C
e
–
P
D

F
/

1
3
3
1
4
2
1
8
2
8
7
4
4
0
0
1
1
5
2
6
0
4
7
7
2
7
4
6
6
8
4
P
D

B
sì
G
tu
e
S
T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

Patricia
Smith
Churchland
SU
apprendimento

Lashley reasoned that if a rat learned

something, such as a route through a
certain maze, and if that information
was stored in a single, punctate location,
then you should be able to extract it by
lesioning the rat’s brain in the right
place. Lashley trained twenty rats on his
maze. Next he removed a different area
of cortex from each animal, and allowed
the rats time to recover. He then retested
each one to see which lesion removed
knowledge of the maze. Lashley discov-
ered that a rat’s knowledge could not
be localized to any single region; it ap-
peared that all of the rats were some-
what impaired and yet somewhat com-
petent–although more extensive tissue
removal produced more serious memory
de½cit.

As improved experimental protocols
later showed, Lashley’s non-localization
conclusion was essentially correct.
There is no such thing as a dedicated
memory organ in the brain; informazione
is not stored on the ½ling cabinet model
at all, but distributed across neurons.
A general understanding of what it
means for information to be distributed
over neurons in a network has emerged
from computer models. The basic idea
is that arti½cial neurons in a network,
by virtue of their connections to other
arti½cial neurons and of the variable
strengths of those connections, can pro-
duce a pattern that represents something
–such as a male face or a female face, O
the face of Churchill. The connection
strengths vary as the arti½cial network
goes through a training phase, during
which it gets feedback about the adequa-
cy of its representations given its input.
But many details of how actual neural
nets–as opposed to computer-simulated
ones–store and distribute information
have not yet been pinned down, and so
computer models and neural experi-
ments are coevolving.

Neuroscientists are trying to under-
stand the structure of learning by using a
variety of research strategies. One strat-
egy consists of tracking down experi-
ence-dependent changes at the level of
the neuron to ½nd out what precisely
i cambiamenti, Quando, and why. Another strate-
gy involves learning on a larger scale:
what happens in behavior and in partic-
ular brain subsystems when there are le-
sions, or during development, or when
the subject performs a memory task
while in a scanner, O, in the case of ex-
perimental animals, when certain genes
are knocked out? At this level of inquiry,
psicologia, neuroscience, and molecu-
lar biology closely interact.

Network-level research aims to strad-
dle the gap between the systems and the
neuronal levels. One challenge is to un-
derstand how distinct local changes in
many different neurons yield a coherent
global, system-level change and a task-
suitable modi½cation of behavior. How
do diverse and far-flung changes in the
brain underlie an improved golf swing
or a better knowledge of quantum
mechanics?

What kinds of experience-dependent
modi½cations occur in the brain? From
one day to the next, the neurons that col-
lectively make me what I am undergo
many structural changes: new branches
can sprout, existing branches can ex-
tend, and new receptor sites for neuro-
chemical signals can come into being.
D'altra parte, pruning could de-
crease branches, and therewith decrease
the number of synaptic connections be-
tween neurons. Or the synapses on re-
maining branches could be shut down
altogether. Or the whole cell might die,
taking with it all the synapses it formerly
supported. Or, ½nally, in certain special
regions, a whole new neuron might be
born and begin to establish synaptic
connections in its region.

Dedalo Inverno 2004

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
D
UN
e
D
UN
R
T
io
C
e
–
P
D

F
/

1
3
3
1
4
2
1
8
2
8
7
4
4
0
0
1
1
5
2
6
0
4
7
7
2
7
4
6
6
8
4
P
D

B
sì
G
tu
e
S
T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

How do
neurons
know?

And that is not all. Repeated high rates

of synaptic ½ring (spiking) will deplete
the neurotransmitter vesicles available
for release, thus constituting a kind of
memory on the order of two to three
seconds. The constituents of particular
neurons, the number of vesicles released
per spike, and the number of transmitter
molecules contained in each vesicle, can
change. And yet, somehow, my skills re-
main much the same, and my autobio-
graphical memories remain intact, even
though my brain is never exactly the
same from day to day, or even from min-
ute to minute.

No ‘bandleader’ neurons exist to en-
sure that diverse changes within neu-
rons and across neuronal populations
are properly orchestrated and collective-
ly reflect the lessons of experience. Nev-
ertheless, several general assumptions
guide research. For convenience, IL
broad range of neuronal modi½ability
can be condensed by referring simply to
the modi½cation of synapses. The deci-
sion to modify synapses can be made
either globally (broadcast widely) O
locally (targeting speci½c synapses). If
made globally, then the signal for change
will be permissive, in effect saying, “You
may change yourself now”–but not dic-
tating exactly where or by how much or
in what direction. If local, the decision
will likely conform to a rule such as this:
If distinct but simultaneous input signals
cause the receiving neuron to respond
with a spike, then strengthen the con-
nection between the input neurons and
the output neurons. On its own, a signal
from one presynaptic (sending) neuron
is unlikely to cause the postsynaptic (Rif-
ceiving) neuron to spike. But if two dis-
tinct presynaptic neurons–perhaps one
from the auditory system and one from
the somatosensory system–connect to
the same postsynaptic neuron at the
same time, then the receiving neuron is

more likely to spike. This joint input ac-
tivity creates a larger postsynaptic effect,
triggering a cascade of events inside the
neuron that strengthens the synapse.
This general arrangement allows for dis-
tinct but associated world events (per esempio.,
blue flower and plenty of nectar) to be
modeled by associated neuronal events.

The nervous system enables animals to
make predictions.7 Unlike plants, ani-
mals can use past correlations between
classes of events (per esempio., between red cher-
ries and a satisfying taste) to judge the
probability of future correlations. A cen-
tral part of learning thus involves com-
puting which speci½c properties predict
the presence of which desirable effects.
We correlate variable rewards with a fea-
ture to some degree of probability, so
good predictions will reflect both the
expected value of the reward and the
probability of the reward’s occurring;
this is the expected utility. Humans and
bees alike, in the normal course of the
business of life, compute expected utili-
ty, and some neuronal details are begin-
ning to emerge to explain how our
brains do this.

To the casual observer, bees seem to
visit flowers for nectar on a willy-nilly
basis. Closer observation, Tuttavia, Rif-
veals that they forage methodically. Not
only do bees tend to remember which
individual flowers they have already vis-
ited, but in a ½eld of mixed flowers with
varying amounts of nectar they also
learn to optimize their foraging strategy,
so that they get the most nectar for the
least effort.

Suppose you stock a small ½eld with
two sets of plastic flowers–yellow and
blue–each with wells in the center into
which precise amounts of sucrose have

7 John Morgan Allman, Evolving Brains (Nuovo
York: Scienti½c American Library, 1999).

Dedalo Inverno 2004

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
D
UN
e
D
UN
R
T
io
C
e
–
P
D

F
/

1
3
3
1
4
2
1
8
2
8
7
4
4
0
0
1
1
5
2
6
0
4
7
7
2
7
4
6
6
8
4
P
D

B
sì
G
tu
e
S
T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

Patricia
Smith
Churchland
SU
apprendimento

been deposited.8 These flowers are ran-
domly distributed around the enclosed
½eld and then baited with measured vol-
umes of ‘nectar’: all blue flowers have
two milliliters; one-third of the yellow
flowers have six milliliters, two-thirds
have none. This sucrose distribution
ensures that the mean value of visiting
a population of blue flowers is the same
as that of visiting the yellow flowers,
though the yellow flowers are more
uncertain than the blues.

After an initial random sampling of
the flowers, the bees quickly fall into a
pattern of going to the blue flowers 85
percent of the time. You can change
their foraging pattern by raising the
mean value of the yellow flowers–for
esempio, by baiting one-third of them
with ten milliliters. The behavior of the
bees displays a kind of trade-off between
the reliability of the source type and the
nectar volume of the source type, con
the bees showing a mild preference for
reliability. What is interesting is this:
depending on the reward pro½le taken
in a sample of visits, the bees revise their
strategy. The bees appear to be calculat-
ing expected utility. How do bees–mere
bees–do this?

In the bee brain there is a neuron,

though itself neither sensory nor motor,
that responds positively to reward. Questo
neuron, called VUMmx1 (‘vum’ for
short), projects very diffusely in the bee
brain, reaching both sensory and motor
regions, as it mediates reinforcement
apprendimento. Using an arti½cial neural net-
lavoro, Read Montague and Peter Dayan
discovered that the activity of vum rep-
resents prediction error–that is, the dif-
ference between ‘the goodies expected’

8 This experiment was done by Leslie Real,
“Animal Choice Behavior and the Evolution of
Cognitive Architecture,” Science (1991): 980–
986.

and ‘the goodies received this time.’9
Vum’s output is the release of a neuro-
modulator that targets a variety of cells,
including those responsible for action
selection. If that neuromodulator also
acts on the synapses connecting the sen-
sory neurons to vum, then the synapses
will get stronger, depending on whether
the vum calculates ‘worse than expect-
ed’ (less neuromodulator) or ‘better
than expected’ (more neuromodulator).
Assuming that the Montague-Dayan
model is correct, then a surprisingly sim-
ple circuit, operating according to a fair-
ly simple weight-modi½cation algo-
rithm, underlies the bee’s adaptability
to foraging conditions.

Dependency relations between phe-
nomena can be very complex. In much
of life, dependencies are conditional and
probabilistic: If I put a fresh worm on
the hook, and if it is early afternoon,
then very probably I will catch a trout here.
As we learn more about the complexities
of the world, we ‘upgrade’ our represen-
tations of dependency relations;10 we
Imparare, Per esempio, that trout are more
likely to be caught when the water is
cool, that shadowy pools are more
promising ½sh havens than sunny pools,
and that talking to the worm, entreating
the trout, or wearing a ‘lucky’ hat makes
no difference. Part of what we call intel-
ligence in humans and other animals is
the capacity to acquire an increasingly
complex understanding of dependency
relations. This allows us to distinguish

9 See Read Montague and Peter Dayan, “Neu-
robiological Modeling,” in William Bechtel,
George Graham, and D. UN. Balota, eds., A Com-
panion to Cognitive Science (Malden, Massa.:
Blackwell, 1998).

10 Clark N. Glymour, The Mind’s Arrows (Camera-
ponte, Massa.: con la stampa, 2001). See also Alison
Gopnik, Andrew N. Meltzoff, and Patricia K.
Kuhl, The Scientist in the Crib (New York: Wil-
liam Morrow & Co., 1999).

Dedalo Inverno 2004

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
D
UN
e
D
UN
R
T
io
C
e
–
P
D

F
/

1
3
3
1
4
2
1
8
2
8
7
4
4
0
0
1
1
5
2
6
0
4
7
7
2
7
4
6
6
8
4
P
D

B
sì
G
tu
e
S
T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

How do
neurons
know?

fortuitous correlations that are not gen-
uinely predictive in the long run (per esempio.,
breaking a tooth on Friday the thir-
teenth) from causal correlations that are
(per esempio., breaking a tooth and chewing hard
candy). This means that we can replace
superstitious hypotheses with those that
pass empirical muster.

Like the bee, humans and other ani-
mals have a reward system that mediates
learning about how the world works.
There are neurons in the mammalian
brain that, like vum, respond to re-
ward.11 They shift their responsiveness
to a stimulus that predicts reward, O
indicates error if the reward is not forth-
coming. These neurons project from a
brainstem structure (the ventral tegmen-
tal area, or ‘vta’) to the frontal cortex,
and release dopamine onto the postsy-
naptic neurons. The dopamine, only
one of the neurochemicals involved
in the reward system, modulates the
excitability of the target neurons to the
neurotransmitters, thus setting up the
conditions for local learning of speci½c
associations.

Reinforcing a behavior by increasing
pleasure and decreasing anxiety and pain
works very ef½ciently. Nevertheless,
such a system can be hijacked by plant-
derived molecules whose behavior mim-
ics the brain’s own reward system neu-
rochemicals. Changes in reward system
pathways occur after administration of
cocaine, nicotine, or opiates, all of which
bind to receptor sites on neurons and are
similar to the brain’s own peptides. IL
precise role in brain function of the large
number of brain peptides is one of neu-
roscience’s continuing conundrums.12

11 See Paul W. Glimcher, Decisions, Uncertainty,
and the Brain (Cambridge, Massa.: con la stampa,
2003).

12 I am grateful to Roger Guillemain for dis-
cussing this point with me.

These discoveries open the door to
understanding the neural organization
underlying prediction. They begin to
forge the explanatory bridge between
experience-dependent changes in single
neurons and experience-dependent
guidance of behavior. And they have be-
gun to expose the neurobiology of addic-
zione. A complementary line of research,
meanwhile, is untangling the mecha-
nisms for predicting what is nasty. Al-
though aversive learning depends upon a
different set of structures and networks
than does reinforcement learning, here
too the critical modi½cations happen
at the level of individual neurons, E
these local modi½cations are coordinat-
ed across neuronal populations and inte-
grated across time.

Within other areas of learning re-
search, comparable explanatory threads
are beginning to tie together the many
levels of nervous system organization.
This research has deepened our under-
standing of working memory (holding
information at the ready during the ab-
sence of relevant stimuli) spatial learn-
ing, autobiographical memory, motor
skills, and logical inference. Granting
the extraordinary research accomplish-
ments in the neuroscience of knowledge,
nevertheless it is vital to realize that
these are still very early days for neuro-
science. Many surprises–and even a rev-
olution or two–are undoubtedly in
store.

Together, neuroscience, psicologia,

embryology, and molecular biology are
teaching us about ourselves as knowers–
about what it is to know, Imparare, remem-
ber, and forget. But not all philosophers
embrace these developments as prog-
ress.13 Some believe that what we call

13 I take it as a sign of the backwardness of aca-
demic philosophy that one of its most esteemed
living practitioners, Jerry Fodor, is widely sup-

Dedalo Inverno 2004

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
D
UN
e
D
UN
R
T
io
C
e
–
P
D

F
/

1
3
3
1
4
2
1
8
2
8
7
4
4
0
0
1
1
5
2
6
0
4
7
7
2
7
4
6
6
8
4
P
D

B
sì
G
tu
e
S
T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

their relative success in predicting and
explaining.

But does all of this mean that there is a
kind of fatal circularity in neuroscience
–that the brain necessarily uses itself to
study itself? Not if you think about it.
The brain I study is seldom my own, Ma
that of other animals or humans, and I
can reliably generalize to my own case.
Neuroepistemology involves many
brains–correcting each other, testing
each other, and building models that can
be rated as better or worse in character-
izing the neural world.

Is there anything left for the philoso-
pher to do? For the neurophilosopher, at
least, questions abound: about the inte-
gration of distinct memory systems, IL
nature of representation, the nature of
reasoning and rationality, how informa-
tion is used to make decisions, what
nervous systems interpret as informa-
zione, and so on. These are questions with
deep roots reaching back to the ancient
Greeks, with ramifying branches ex-
tending throughout the history and phi-
losophy of Western thought. They are
questions where experiment and theo-
retical insight must jointly conspire,
where creativity in experimental design
and creativity in theoretical speculation
must egg each other on to unforeseen
discoveries.14

14 Many thanks to Ed McAmis and Paul
Churchland for their ideas and revisions.

Patricia
Smith
Churchland
SU
apprendimento

external reality is naught but an idea cre-
ated in a nonphysical mind, a mind that
can be understood only through intro-
spection and reflection. To these philos-
ophers, developments in cognitive neu-
roscience seem, at best, irrelevant.

The element of truth in these philoso-
phers’ approach is their hunch that the
mind is not just a passive canvas on
which reality paints. Infatti, we know
that brains are continually organizing,
structuring, extracting, and creating. As
a central part of their predictive func-
zioni, nervous systems are rigged to
make a coherent story of whatever input
they get. ‘Coherencing,’ as I call it,
sometimes entails seeing a fragment as
a whole, or a contour where none exists;
sometimes it involves predicting the im-
minent perception of an object as yet
unperceived. As a result of learning,
brains come to recognize a stimulus as
indicating the onset of meningitis in a
child, or an eclipse of the Sun by the
Earth’s shadow. Such knowledge de-
pends upon stacks upon stacks of neural
networks. There is no apprehending the
nature of reality except via brains, E
via the theories and artifacts that brains
devise and interpret.

From this it does not follow, Tuttavia,
that reality is only a mind-created idea. It
means, Piuttosto, that our brains have to
keep plugging along, trying to devise
hypotheses that more accurately map
the causal structure of reality. We build
the next generation of theories upon the
scaffolding–or the ruins–of the last.
How do we know whether our hypothe-
ses are increasingly adequate? Only by

ported for the following conviction: “If you
want to know about the mind, study the mind
–not the brain, and certainly not the genes”
(Times Literary Supplement, 16 May 2003, 1–2).
If philosophy is to have a future, it will have
to do better than that.

Dedalo Inverno 2004

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
D
UN
e
D
UN
R
T
io
C
e
–
P
D

F
/

1
3
3
1
4
2
1
8
2
8
7
4
4
0
0
1
1
5
2
6
0
4
7
7
2
7
4
6
6
8
4
P
D

B
sì
G
tu
e
S
T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3
Scarica il pdf