Professional Documents
Culture Documents
Ro^Jtrb fl
Vehicles
lr
db
Bradford Books
Vehicles
ExDerimentsin
SyhtheticPsychology
Valentino Braitenberg
Sixth printing,1998
First MIT Presspaperback edition, 1986
@ 1984 by The MassachusettsInstitute of Technology
All rights reserved.No part of this book may be reproduced
in any form by any electronic or mechanicalmeans(including
photocopying, recording or information storageand retrieval)
without permissionin writing from the publisher.
Jacket artwork and portfolio of drawings
by Maciek Albrecht:
Copyright @ 1984 by Maciek Albrecht
All rights reserved.This arnvork may not be
reproducedin any form or by any means,electronic
or mechanical,including photocopying or by any
other means,without permission in writing from
Maciek Albrecht.
Technicaldrawings by Ladina Ribi and Claudia
Manin-Schuben.
This book was set in Linotron eoz Sabonby Achorn
Graphic Servicesand printed and bound
in the United Statesof America
Library of CongressCataloging in Publication Data
Braitenberg,Valentino.
Vehicles,experimentsin syntheticpsychology.
"Bradford books"-Series t.p.
Bibliography: p.
r. Neurophysiology. z. Psychology,Physiological.
r. Tide.
Gontents
ix
GemingAround 3
Vehicle z
Vehicle 3
Love ro
Vehicle 4
Vehicle 5
Logic
Vehicle 6
Selection,the ImpersonalEngineer z6
Vehicle 7
Concepts 29
Vehicle 8
Vehicle 9
Shapes 4j
Vehicle ro
Getting Ideas to
Vehicle r r
Vehicle rz
Trains of Thought 6z
20
Vehicle 13 Foresight 70
Vehicle 14 Egotism and Optimism 8o
vii I CoNTENTS
Fmewod
I FOREWORD
Vehicles
Inhoduction
Z I VEHICLES
to you, if you arepreparedto follow me not througha world of real
brains but through a toy world that we will createtogether.
We will talk only about machineswith very simpleinternal structures, too simplein fact to be interestingfrom the point of view of
mechanicalor electricalengineering.Interestarises,rather, when
we look at thesemachinesor vehiclesas if they were animalsin a
natural environment.We will be tempted,then, to usepsychological languagein describingtheir behavior.And yet we know very
well that there is nothing in thesevehiclesthat we havenor put in
ourselves.This will be an interestingeducationalgame.
Our vehiclesmay move in water by jet propulsion.Or you may
prefer to imaginethem moving somewherebeirveengalaxies,with
negligible gravitational pull. Remember,however, that their jets
must expel matter in order to function at all, and this implies replenishmentof the food storeswithin the vehicles,which might be a
problem berweengalaxies.This suggestsvehiclesmoving on the
surfaceof the earth through an agriculturallandscapewhere they
have good support and can easily find the food or fuel they need.
(Indeed the first few chaptershere conjure up imagesof vehicles
swimming around in the water, while later what comesto mind are
little carts moving on hard surfaces.This is no accident,if the
evolution of vehiclesr to 14 in any way reflectsthe evolution of
animal species.)
It does not matter. Get used to a way of thinking in which the
hardware of the realizationof an idea is much lessimportant than
the idea itself. Norbert Wiener was emphaticabout this when he
formulated the title of his famous book: Cybernetics,or Control
and Communication in Animals and Machines.
Vehide 1
Getting Around
Figure r
Vehicle r, the simplest vehicle.The speedof the motor (rectangularbox at
the tail end) is controlled by a sensor (half circle on a stalk, at the front
end). Motion is always forward, in the direction of the arrow, except for
perturbations.
I Getting Around
Vehicle2
Figure z
otherwiselike Vehicler. The
Vehiclez, with two motorsandnvo sensors;
differ in a, b, and c.
connections
Figure 3
Vehiclesza andzb in the vicinity of a source(circlewith raysemanating
from it). Vehiclezb orientstowardthe source,za awayfrom it.
9 | Fearany' Aggression
the vehiclewill turn
motor will work harder.And as a consequence
away from the source.
Now let us try the other schemeof sensory-motorconnections,
(b) in figure 3. No changeif the sourceis straight ahead.If it is to
one side,however,we notice a differencewith resPectto Vehicle za.
Vehicle zb will turn toward the sourceand eventuallyhit it. There
is no escaping:as long as zb staysin the vicinity of the source'no
matter how it stumblesand hesitates,it will hit the sourcefrontally
in the end. Only in the unlikely casethat a strongperturbationin its
coursemakesit turn exactly away from the source,and no further
perturbation occurs,can it escapeits fate.
Let Vehiclesza and zb move around in their world for a while
and watch them. Their charactersare quite opposite.Both DIsLIKE
sources.But za becomesrestlessin their vicinity and tendsto avoid
them, escapinguntil it safelyreachesa placewhere the influenceof
the sourceis scarcelyfelt. Vehicleza is a cowARD' you would say.
Not so Vehiclezb. It, too, is excitedby the presenceof sources,but
resolutelyturns toward them and hits them with high velocity,asif
it wanted to destroythem. Vehicle zb is AccREsslvn,obviously.
Vehicle3
Loue
\
\
I
/
Figurc 4
Vehicle3, with inhibitoryinlluenceof the sensors
on the motorc.
rz I VEHICLEI
source.The vehiclewith crossedconnecrions(figure4b) for analogous reasonswill come to rest facing away from the sourceand
may not staytherevery long, sincea slightperturbationcould cause
it to drift away from the source.This would lessenthe source's
inhibitory influence,causingthe vehicleto speedup moreand more
as it gets away.
You will haveno difficulty giving namesto this sort of behavior.
ThesevehiclesLIKEthe source,you will say, but in differentways.
Vehicle 3a LovEs it in a permanentway, stayingcloseby in quiet
admiration from the time it spots the sourceto a[ future time.
Vehicle3b, on the other hand, is an ExpLoRER.
It likes the nearby
sourceall right, but keepsan eyeopen for other, perhapsstronger
sources,which it will sail to, givena chance,in order to find a more
permanentand gratifying appeasement.
But this is not yet the full developmenrof Vehiclet. Ifle are now
readyto make a more completemodelusingall the behavioraltraits
at our disposal.Call it Vehicle 3c. \UUe
give it not just one pair of
sensorsbut four pairs, tuned to different qualitiesof the environment, say light, temperature,oxygenconcentration,and amount of
organic matter (figure 5). Now we connectthe first pair to the
motors with uncrossedexcitatoryconnections,asin Vehicleza, the
secondpair with crossedexcitatoryconnections,as in Vehiclezb,
and the third and the fourth pairs with inhibitory connections,
crossedand uncrossed,as in Vehicles3b and 3a.
This is now'a vehiclewith really interestingbehavior.It dislikes
high temperature,turns awayfrom hot places,and at the sametime
seemsto dislike light bulbs with evengreaterpassion,sinceit turns
toward them and destroysthem. on the other hand it definitely
seemsto prefer a well-oxygenatedenvironmentand one containing
many organic molecules,sinceit spendsmuch of its time in such
places.But it is in the habit of moving elsewherewhenthe supplyof
eitherorganicmatter or (especially)
oxygenis low. You cannothelp
admitting that Vehicle 3c has a systemof veruns, and, come to
Figure 5
A multisensorial
vehicleof brand3c.
14 | VEHICLE3
think of it, KNovLEDcn,since some of the habits it has, like deasif the vehicle
stroying light bulbs,may look quite knowledgeable,
knows that light bulbs tend to heat up the environmentand consequently make it uncomfortableto live in. It alsolooks asif it knows
about the possibility of making energyout of oxygen and organic
matter becauseit prefersplaceswhere thesetwo commoditiesare
available.
But, you will say,this is ridiculous: knowledgeimplies a flow of
information from the environment into a living being or at least
into somethinglike a living being.Therewas no suchtransmission
of information here.We were iust playingwith sensors,motors, and
connections:the propertiesthat happenedto emergemay look like
knowledgebut really are not. lU?eshouldbe carefulwith suchwords.
You are right. I[e will explain in a later chapter (on Vehicle 5)
how knowledge may enter a systemof connections.And we will
introduce an alternativeway of incorporatingknowledgeinto the
systemin our chapteron Vehicle7. In any case,onceknowledgeis
incorporated,the resultingvehiclemay look and behavequite like
our Vehicle 7c.
Meanwhile I invite you to considerthe enormouswealth of different propertiesthat we may give Vehicle3c by choosingvarious
sensorsand various combinationsof crossedand uncrossed,excitatory and inhibitory, connections.
If you considerthe possibilityof strongandweak influencesfrom
the sensorsto the motors, you realizethat the variety becomeseven
greater.The vehiclemay not care much about light but care very
much about temperature.Its senseof smellmay be much keenerfor
organic matter than it is for oxygenor viceversa.And theremay be
many more than just four pairs of sensorsand four sensoryqualities: the vehiclesmay be equipped with all sorts of shrewd detectors of energy and of chemicals.But this is best discussedin
connectionwith a new idea incorporatedin the vehiclesof the next
chapter.
Vehicle4
Valuesand SpecialTastes
16 | V E HI CL E +
Figure 5
A nonlinear dependence of the speed of the motor V on the intensity of
stimulation I, with a maximum for a certain intensity.
17 | Valuesand SpecialTastes
on one side or the other of the maximum describingthe sensorymotor dependence(figure Z). Vehicle 4a might like one sort of
stimuluswhen it is weak but not when it is too strong; it might like
anotherstimulusbetter the strongerit becomes.It might turn away
from a weak smell and destroythe sourceof a strong one. It might
visit in alternationa sourceof smelland a sourceof sound,turning
away from both with a changeof temperature.
Watchingvehiclesof brand 4ain a landscapeof sources,you will
be delightedby their complicatedtraiectories.And I am sure you
will feel that their motives and tastesare much too varied and
intricateto be understoodby the observer.Thesevehicles,you will
say, are governedby rNsrtNcrs of various sorts and, alas,we iust
don't know how Nature managesto embody instinctsinto a piece
of brain.
You forget, of course, that we have ourselvesdesignedthese
vehicles.
But instincts are a lowly sort of behavior anyway. We can do
better.Let us improve on type 4aby adding a new sort of connection benrveensensorsand motors. This time the influenceof the
sensoron the motor is no longersmooth; there are definitebreaks.
There might be a range of intensitiesof sensorystimulation for
which the motor is not activatedat all and then, under stronger(o,^lil|,1
stimuli,the motors are running at full speed.Or else,theremight be .rF ,^1.,*ft
smoothchangesof motor activationfor certain ranges,with abrupt
changesin between.A very lifelike paffern would be: no activation
up to a threshold value of the stimulus, and increasingactivation
beyondthe threshold,startingwith a certainfixed minimum (figure
8). You are by now experiencedin the art of creativeinventionand
will haveno difficulty dreamingup more schemesof this sort.
In a way thesenew vehicles,which we call 4b, are alreadycontainedin the vast classof vehicles4?, sinceabruptnessof behavior
can of coursebe simulatedwith any degreeof approximation by
that are in reality, mathematicallyspeakfunctional dependences
\\/
\;\
//
\l/
/
|
./'.
-\r( s . ) - \
.r\-7:
-.-
-'
\'---?
/t\
/
,/\
tlz
t'
l :Ai
..\-/r.
/l\
\--/
/ -a.
I
/
\)t
t/
\\_/ ----_
.,.2
Figure 7
sources.
Traiectoriesof vehiclesof brand4a aroundor benveen
/
)
\
19 | Valuesand SpecialTastes
Figure 8
Variousbizarrekindsof dependence
of the speedof the motor (ordinate)
on the intensityof stimulation(abscissa)
in Vehicle4b.
ing, continuous.Moreover, if friction plays a role, as we have al'
readydecidedit should,thresholdsin motor activationwould ensue
naturally: the vehiclewill staft moving only when the force exerted
by.the motor exceedsa certain value, sufficientto overcomethe
initial friction.
Whatever their origin, thresholds in some behavior patterns
make a lot of differencein the eyeof the observer.Thesecreatures,
'!7hen
you
the observerwould say, ponder over their DEcIsIoNs.
come close to them with a lure, it takes them some time to get
going. Yet once they havedecided,they can act quite quickly. Th.y
do indeed seemto act in a spontaneousway: none of this passive
being attracted one way or the other that was so obvious in the
vehiclesof the more lowly types.You would almostbe temptedto
say:where decisionsare beingmade,theremust be a wIn to make
them. Why not? For all we know, this is not the worst criterion for
establishingthe existenceof free will.
Vehicle5
Logic
\ ,/
x_
At this point we are ready.tomakea fundamentaldiscovery. We have gatheredevidencefor what I would like to call the
"law of uphill analysisand downhill invention." What I mean is
this. It is pleasurableand easy to createlittle machinesthat do
certain tricks. It is also quite easyto observethe full repertoireof
behavior of thesemachines-even if it goesbeyond what we had
originally planned,as it often does.But it is much more difficult to
start from the outside and to try to guessinternal structureiust
from the observationof behavior.It is actuallyimpossiblein theory
to determineexactly what the hidden mechanismis without openwith
ing the box, sincethere are alwaysmany differentmechanisms
identical behavior.Quite apart from this, analysisis more difficult
than invention in the sensein which, generally,induction takes
more time to perform than deduction: in induction one has to
searchfor the wxy, whereasin deductionone follows a straightforward path.
A psychologicalconsequenceof this is the following: when we
analyzea mechanism,we tend to overestimateits complexity. In the
uphill processof analysis,a givendegreeof complexityoffersmore
resistanceto the workings of our mind than it would if we encoun-
zr I Logic
teredit downhill, in the processof invention.We havealreadyseen
this happenwhen the observerof Vehicle4b conjecturedthat the
vehicledoessome thinking beforeit reachesa decision,suggesting
complicatedinternal processeswhere in reality there was nothing
but a thresholddevicewaiting for sufficientactivation.The patterns
of behaviordescribedin the vehiclesof type 4aundoubtedlysuggest
much more complicatedmachinerythan that which was actually
usedin designingthem.
We may now take pleasurein this and createsimple"brains" for
our vehicles,which will indeed(asexperienceshows)tax the mind
of eventhe most playful analyst.All we have to do is introduce
specialelements,calledthresholddevices,which will be eitherinterposedberweensensorsand motors or connectedto eachother in
complexesthat receivesomeinput from the sensorsand give some
output to the motors.
The individual thresholddeviceis of the simplestsort: it givesno
output if its input line carriesa signalbelow the threshold,and it
givesfull output beyond the threshold.We will also use anorher
variety giving output all the time unlessthe input carriesa signal
above the threshold. Each of thesedevicesis fitted with a knob
which may be turned to set the threshold,so that the input would
becomeeffectivewith one, two, or any specifiednumber of input
activationunits. (The word thresholdof courseimplies that, for a
given threshold value, any input strongerthan the one specified
would also be effective.)
We arenot limited to the typesof connectionsthrough which the
thresholddevicesactivateeachother.we can alsouseanotherkind,
call them "inhibitoryr" which counteractthe activationthat comes
from other sources(figure 9).
In order to make a brain out of thresholddevices,we may connectthem togetherone to one, or many to one, or one to many, or
manyto one and one to many, in whicheverway we like.'Sfhenyou
are designingbrains, it is important for you to know that in one of
z z I V E HI CL E 5
Figure 9
How threshold devices act on each other. Explanation of symbols: The
circles stand for threshold devices. The L-shaped fiber berween B and C
stands for inhibition; the penetrating fiber from A to C means activation.
Each active element contributes one unit of activation to the element
(threshold device) to which
threshold device becomes active when the activation reaches at least the
threshold value indicated within the circle. An inhibitory connection from
an active element subtracts r from the sum of all the units of activation
reaching the same target element. A negative threshold (or threshold o)
implies activity in the absence of external activation. Such an element can
5\,rt)
Q, n,i -)
-Tfn
\{
1ftU
+" '.q,hh ,-,6afr'i"srartrs
zt
I Logic
Figure roa
A network that gives a signal when a burst of 3 pulsespresentsitself,
precededand followpd by a pause.
AcArurr
*"dtr l-n
4-
Figure rob
A nenvorkof thresholddevicesthat emitsa pulsefor everythird pulsein a
row in the input.
But Vehicle 5 can do much more than that. It can count (figure
ro). It may associateonly with groups of four vehicles,not more
and not less,to make a party of five. Or it may visit every tenth
sourceit encounterson its way. Or it may turn awayfrom a vehicle
whosenumberof sensorsis a multiple of seven,implying that such
vehiclesbring bad luck. In some way, it 'tseemsto operate with
'
NUMBERs
.
tztlugr q" hf(/p
^,n\nj
Lan f re Iz /a\1*",
If you fit such a vehicle with a very large roilb., oI shrewdly
connectedthresholddevices,you may get it to play a passablegame
of chess.Or you may make it solve puzzlesin rocrc or prove
theoremsin euclideangeometry.You realizewhat I am driving at:
with enoughthresholddevicesit can do anything a computer can
do, and computerscan be made to do almost everything.
But where is the memory, some of you will ask, realizing that
p4
24 | VEHICLE 5
most of the activitiesof a digital computerconsistof putting data
into memory, taking the data out again to perform somecalculation, putting the resultsback into the memory, and so forth. The
answer: there is room for memory in a network of thresholddevices,if it is largeenough.Imaginea thresholddeviceconnectedto
'$(Ihen
a sensor for red light.
it is activated by the red light, it
activatesanotherthreshold devicewhich in turn is connectedback
to the first device.Once a red light is sighted,the nvo deviceswill
activateone anotherforever.Take a wire from the output of one of
the two thresholddevicesand connectit to a bell: the ringing of the
bell then signalsthe fact that at sometime in the past this particular
vehicle sailedin the vicinity of a sourceof red light.
This is an elementarysort of unuonv. It is not difficult to understand how out of such elementarymemory stores (consistingof
reciprocallyconnectedthresholddevices)complexmemoriescanbe
synthesized,with the possibility of storing extremely complex
events.But there is a limit to the quaatity of facts the vehiclecan
store this way. For instance,when storing numbers,if the vehicle
has a bank of ten elementarymemory devices,it cannot fit any
number that has more than ten digits (in binary notation), since
qlch elementarydevicecan at most rememberone digit by being
active or inactive ("one bit of information").
There is a trick that can be used by our brand 5 vehiclesto
overcomethe intrinsic limitation of their storagecapacity.Imagine
a vehicleinvolvedin a calculationin which numbersoccur that are
much larger than the number of parts in the vehicle'sown interior.
You might think that such a task would be forever beyond the
comprehensionof that particular vehicle.Not so if we employ the
following strategy. Let's transfer our vehicle to a large, sandy
beach. The vehicle can crawl on the beach,leaving marks in the
sand indicating the successionof digits in the large numbersthat
emerge from its calculations.Then it can crawl back, following
zS I Logic
its own track, to rgad off the digits and put them back into the
calculation. A
T/\
)a g t,
Vehicle6
Selection,the Impersonal
Engineer
27 | Selection
course we will not pick up vehicles that have fallen on the floor
becausethey have proved their own inability to cope with the environment. We will be careful to produce vehicles at a pace that
roughly matches the rate at which vehicles fall off the table, to
prevent the race from dying out, on one hand, and to prevent the
table from becoming unduly crowded, on the other.
28 I VEHICLE6
unwittingly introduced into the wiring, though improbable,do becomequite widespreadin the long run.
. .'
This story is quite old and goesby the,nameof Darwinian evolution. Many peopledon't like the idea tliat everythingbeautifuland
marvelousin organic nature shoul{,,6edue to the simplecooperation of reproduction, errors, and/election. This is no problem for
us. We have convinced oursefvi:sthat beautiful, marvelous,and
shrewdmachinescan be mailCout of inorganicmamerby this simple trick. Moreover, we alfiady know that analysisis much more
difficult than synthesis.,.&hetethere has been no consciousengineeringat all, as i
caseof our type 6 vehicles,analysiswill
necessarilyproducldthe feeling of a mysteriou.ssupernaturalhand
guiding the
. We can imagine that in most casesour analysisof brains,/rypr 6 vehicleswould fail altogether:the wiring that
produces
behavior may be so complicatedand involved that
we will7,{everbe able to isolatea simplescheme.And yet it works.
Vehicle7
Concepts
'\U7e
have already usedthe word knowledge,evenif in a
somewhatfacetiousway, when we discussed
the propertiesof Vehicle 3. And we haveiust observedhow a processakin to Darwinian
evolution may incorporate knowledgeinto machinesin a mysterious way, though it is not immediatelyobviousthrough what channel the knowledge (about the dangersconnectedwith a cliff)
enteredthe "brain" or in what form it is containedthere. In both
caseswe are referring to fixed, inborn knowledge that, whether
right or wrong, belongsto the individual vehiclefor better or for
worse.This is fine for a set environmentbut may be catastrophic
when the conditionschange.Therefore,in a preciousvehiclethat
we love, we should build in mechanisrns
of adaptationto make it
more flexible.Not only will our vehiclethen be preparedro meet
catastrophiceventsbut it will also be readyto cope with a greater
variety of situations and thus be less confined ro a particular
environment.
We proceedas follows. First, we buy a roll of a specialwire,
calledMnemotrix, which has the following interestingproperry: its
resistanceis at first very high and stayshigh unlessthe two componentsthat it connectsare at the sametime traversedby an electric
3o I VEHICLEz
current. When this happens,the resistance
of Mnemotrix decreases
and remains low for a while, little by little returning to its initial
value.* Now let's put a piece of Mnemotrix between any two
threshold devicesof a fairly complicatedvehicleof type 5. This is a
lot of wiring, but the effect is not great at firsr, due to the high
resistanceof Mnemotrix. Very little current will spread from an
active component to all the other componentsto which it is
connected.
As the vehicle(which is now type 7) movesaround and experiencesvarious situationsin its environment,someof its Mnemotrix
connectionswill changetheir strength.Supposeaggressive
vehicles
in that particular environmentare often paintedred. Then the sensor for red in our type 7 vehiclewill often be activatedtogetherwith
the threshold devicethat respondsto aggressive
behavior,and the
Mnemotrix wire connectingthe two will have its resistancedecreasedso often that it will not have time to return to its initial
value. The consequenceis obvious: every time the vehicle senses
red, the whole setof movementswith which it normally respondsto
aggressivebehaviorwill be activated.Soour vehiclewill turn away
from its dangerousfellow. The enhancedconnection between the
components representswhat philosopherscall AssocrATron,the
associationof the color red with aggression.More generally,we
may say a new coNcEpr has arisen in the vehicle: wheneveran
aggressivevehicleis around, evenif it is blue or green,our type 7
vehicle will "see red." As far as we are concerned,this can mean
* I don't care if the electriciansshudder.They know very well
that evenif Mnemotrix
is not available commerciallyas a wire, it can be simulatedby a simple circuit. And
they also know that suchthings exist in animals' brains.If you want a fairly realistic
explanation of Mnemotrix wire, think of a material that changesits conductanceas
a function of temperature: the current heats the two components connectedby
Mnemotrix, and the temperaturechangeat the two endsof the connectioninduces
the change in resistance.
Jt
I Concepts
only: the vehicle does some of the things it did previously only \-\
\
when it was confrontedwith the color red.
This processof translating things that happen together in the rl.
-_I-.
J
environmentinto "complexes" of activity within the vehiclesis of \ 5 - <i
suchgreatimportancethat we ought to familiarizeourselveswith it
some more. One consequence,
we have already seen,is concept
formation. when it happensbetweendifferent categoriesof things
(suchas red color and aggression),
we prefer to call it association.
But it may happen within a single category, say smell, when a
number of chemicalsdissolvedin the air are frequently perceived
together,suchas burned plastic,lubricatingfluid, and battery acid,
which are set free when a vehicleis wrecked.So it is iustified for
survivingvehiclesto storethe "smell of death" in order to be able,
later on, to identify dangerousregionsof their environment.This is
done by the formation of a new olfactory concept.
Visual concepts may be formed in a similar manner. Tlie
straightnessof a line in differenr parts of the visual field, for example, may come to signify the dangerouscliff at rhe side of the.
table. And the movementof many obiectsin different directions
may cometo representthe concept"region crowded with vehicles."
But visualconceptscan be treatedmoreefficientlylater on when we
provide our vehicleswith the a priori categoryof space.For now,
we should explore some of the philosophicalimplications of the
processof conceptformation.
Let philosopherswatch a breedof rype 7 vehiclesand let
$Sm
speculate
about the vehicles'behavior.One philosophersaysi4his
is all very well, but learningto recognizesituationsthat are of some
importanceis a fairly trivial performance,especiallyif it is donethe
hard way, by reward and punishment.It would be a different matter altogether if these vehicles could form their own conceptsin
quiet meditation, without an external tutor telling them what is
important. But they never will, becauseabstraction is one of the
powersthat is unique to the human mind. \\
,z I VEHICLEz
But look, saysanother philosopher,I iust watchedan ABsrRAcTIoN beingmadeby one of thesecreatures.It wasmovingaround in
a crowd of peaceful,unpainted gray vehicleswhen it met a vehicle
painted red that proved to be aggressive;
then it met a greenvehicle
that also proved to be aggressive.When my vehiclemet another
painted fellow, this one painted blue, it immediatelythought that
too. And it turned awayin a hurry. This is a
this one was aggressive
true abstraction,the conc'ept'ofcolor replacingthe individual colors red and greenof the original experience.Or if you wish, we can
has taken placefrom particular colors
say that a cENERALIZATIoN
indicating dangerto the generaldangersignal"color."
Sureenough,saysthe third philosopher,but that is not difficult
to explain either. It has somethingto do with the way colors are
representedby the activity of the electronicparts in the gadget.
Undoubtedly in all the messof wires there will be one wire that
signifies"gray" as the evenmixture of all colors.Then theremight
well be one that signifies"not gt?yr" and that one was activewhen
the red vehicleappeared.So the "not gray" wire had the strongest
and this was learned.No wonder
correlation with aggressiveness,
this "not gray" wire functioned as a dangersignalwhen the blue
aggressorarrived.
All right, saysthe fourth philosopher,but nobody in his right
mind ever suspectedanything more mysteriousbehind the "faculty
of generalization."
Fine, says I, as long as you admit it.
VehicleI
Figure rr
Vehicle8 with a lenseye.
35 | Space,Things,and Mouernents
environmentsand environmentsthat contain lumps of matter,
things that move and orderedstructures.
Build yourselfan array of thresholddevices,eachconnecredro a
group of neighboring photocells,say four of them arrangedin a
square(figurerz). Now aslong asrhevehicleis surroundedby little
insignificantobiectsor by obiectsquite far away,all of the photocellsmight seeiust a few of thesethings,all in more or lessthe same
numbers. Consequently,the photocells will all become active
roughlyto the samedegree.Evenif somephotocellaccidentallysees
a few more things than its neighborsand consequentlygivesa little
more output, the effect will probably be averagedout by the
thresholddevices,which alwaysadd the ourput of four neighboring
photocells.But when a largerobiectappearsin the neighborhoodof
our vehicle,it will be seenby one or more groupsof photocellsthat
are all connectedto the samethreshold device.This devicewill be
activatedmuch more strongly than the others and thus will function as an obiect detector,of inestimablevalue for the vehicle.
It might be even more useful to construct a set of mouernent
detectorsconnectedwith the array of photocells(figurer3). Put the
output of eachphotocellinto a delay,a devicethat givesoff a signal
a little while after it has receivedone. Nothing's easierthan that.
A sluggishthreshold device will do. Now make a new array of
thresholddevices.Eachis connectedwith one photocell via a delay
device,and with anotherneighboringphotocelllocatedto the right
directly, without a delay device.Thesethreshold devicesbecome
activeonly when they receivea signal from both channels.Every
time a bright object movesby from left to right, it will elicit a signal
in onephotocell,which will be storedfor a short while in the delay.
By the time the obiect elicits a signalin the neighboringphotocell,
the delaywill give off its signalaswell so the rwo signalswill hit the
movementdetector-thresholddeviceat the sametime, making it
active.obviously, a spot moving in the oppositedirectionwill not
havethe sameeffect becauseit will hit the fast thresholddevicefirst
Figure rz
An obiectdetector.Eachof the thresholddevices
on the right responds
only when four neighboring sensors arranged in a square are active
together.
Figure 13
A set of movement detectors (C) for movement from right to left. The
threshold devices C become active when they receive input directly from
the sensor F to the left, and at the same time receive input indirectly, via a
delay element D, from the neighboring sensor to the righr
17 | Space,Things,and Mouernents
and the sluggishone afterward-so their output will not coincideat
the next level.Thus our movementdetectorsare directional.
We can of coursemake different setsof movementdetectorsfor
different directionsso that no movementwill escapethe attention
'S7e
can also make them for various velocities,or
of our vehicle.
evenfor objectsof various sizes.In order to do this, we first make
an array of obiect detectors,as in figure rz, and connect their
outputs in pairs to the movement detectors.Only movementof
obfectsof a certain size, defined by the wiring of the individual
obfect detectors,will elicit activity in the movementdetector.We
may also proceedthe other way around.First we make an anay of
movgmentdetectors,all tlned to movementof the samevelocityin
the samedirection.Then we take the output of setsof neighboring
movementdetectorsand connect each set to a threshold device,
which then actsas an obiect detector.But this obiect detectorsees
an object only as a set of points, all moving in the samedirection.
This, by the way, is how we humansseecertainobjectstoo-such
as a cuttlefishmoving on the sandy ocean floor, no maffer how
good the mimicry of the beast.
Another well-known way to make gooduseof an arrayof photocellsis what is often called lateral inhibition (figure r4). Make an
arrayof thresholddevicesbehind the array of photocells.Connect
them one-to-oneto the photocells,so that eachwill be activatedby
light in the correspondingposition. Now introduce lateral inhibition: let eachactivethresholddeviceput a brake on the activity of
its neighbors,so that the more it is activated,the more its neighbors
are inhibited.You can easilyseethat therewill be an unevenmatch
benveenneighboringthresholddevicesreceivingdifferent amounts
of excitation:the one more strongly excitedwill put the other one
completelyout of business.Thus, insteadof getting a continuous
distribution of activity reflectingall the shadesof the environment
seenby the photocells,you will get a representationof isolated
bright spots.Only in the caseof an entirely uniform illumination
38 I VEHICLE8
Figure 14
Fivethresholddevices,
excitedby thatmanysensors,
eachconnected
to its
neighbors
by inhibitoryconnections.
Uniformexcitationof thewholeset
will besubdued
by theinhibitoryinteractions,
whileisolatedspotsof excitition will standout.
will all the thresholddevicesstay at the samelevel (althoughthere
are difficultiesat the borders of the array).But in the caseof uniform illumination, the thresholddeviceswill also inhibit eachother
by the sameamount. Thus uniformiry will be weakly represented,
which is all right, for uniformity is uninteresting.
It is quite clearthat thesetricks, and a numberof other tricks that
you might invent, are only possiblewhen there is an orderly representation of the "sensory space" somewherein the body of the
vehicle.This neednot be z-dimensionalvisual space,as in the examplesjust discussed.
It may be 3-dimensionaltactile space;we can
representinternally,in a 3-dimensionalarray,all the points that the
vehicletouchesby meansof a jointed arm carrying a tactile sensor.
's7e
can also represent3-dimensionalvisual space,if we passrhe
signals from two eyesthrough a devicethat performs the sort of
computationknown as "stereoscopicvision" in human psychology.
'we
can invent all sorts of bizarre internal spaceswhich we might
useto file in a convenientway the information reachingthe vehicle.
Two-dimensionalvisual spacecombinedwith one temporaldimension may lead to a representationof all the images,past and pres-
3g I Space,Things,and Mouements
ent, in a 3-dimensionalspatial arraywithin the vehicle.Inspiredby
someof the things that are known about animal brains,we could
also invent a 3-dimensionalarrayfor the filing of acousticinformation, with one dimensionrepresenting
the frequency,the secondthe
intensity,and the third the phaseof the acousticsignals.
Curiously,when we constructinternal spacesfor vehicles,we are
not even confined by the 3-dimensionalityof familiar spacethat
seemsto limit our immediateintuitive understanding.It is difficult
to imaginesolids of more than 3 dimensions,say a 4-dimensional
cubeor a 5-dimensionalsphere.In fact, when we think of an ordinary 3-dimensionalcube,we tend to imaginesomethinglike a box
with g squaresides.If we want to imaginea 4-dimensionalcube,we
noticethat the sideswould haveto intersect.But we cannotpicture
this, so we give up.
On the other hand, it is quite easy to imagine or to draw
networks of more than 3 dimensions(figure r5). The drawing
shows spheres connected by wires. The nenryork is truly 4-dimensional,sincein order to specifythe coordinatesof one of the
balls (or the path that leadsfrom one ball to a certainother ball),
you haveto indicatehow many stepsto move in directionsx, y, z,
and w. If you disregarddistanceand angleson the drawing (you
can't keep them equal on a projection even in the caseof a 3dimensionalnet), and if you imaginethe net continuedad infinitum
in all 4 directions,the network will look the sameno matterwhich
ball you sit on or in which of the 4 directionsyou look. Now, you
could evenbuild the nenvork, or a piece of it, out of spheresand
wires: you would be able to hold in your hands a structurethat is
intrinsically 4-dimensional, though of course collapsed ("proiected")into the 3 dimensionsof spacein which your handsmove.
(An architect similarly collapseshis buildings into the z-dimensional spaceof his drawing board.) You could even sit on your
nerwork and squashit into a z-dimensionalfelt. It would not matter. A louse finding its way along the wires would still notice the
4 0 | V E HI CL E 8
Figure 15
A four-dimensional
cube.Eachedgeis markedby threeblackdotson a
line, connected by a wire.
4r I Space,Things,and Mouements
ceftain distancein a certain direction, and then again in another
direction.If the placewherethe vehiclewas beforehad somefavorable connotations,it might want to go back. tUfill it move back
exactlythe way it came,or will it choosethe diagonal,which is the
quickestway to get there?If it has an internal representationof.z'
dimensionaleuclideangeometry (that is, if it has z-dimensional
spacebuilt in a priori), it will head directly toward the goal.
Now this internal representationof spaceis somethingthat we
could very easily wire into the network within the vehicle.Just
imagine a z-dimensional sheet made of a material which has
everywherethe sameconductancevaluefor electriccurrents.This is
definedas the current (in amperes)divided by the voltageapplied
(in volts) for a wire of a certainthicknessand a certainlength.Now
if we apply a voltage differencebetweent'wo points on the sheet,
the current that flows through the materialis strongest(the current
area,is highest)alonga straight
density,current per cross-sectional
points.
If
we let one of the nvo points
the
two
line connecting
representthe place where the vehicle is and the other point the
placewhereit wants to go, we caneasilyconstructa devicethat will
determinethe best coursefor the vehicleby way of a simplemeasurementof current densityin different directionson the sheet.
So we would conclude that Vehicle 8 does have the a priori
conceptof z-dimensionalspace.Could Vehicle 8 embodythat of 3and 4-dimensionalspaceas well? To wire an internal representation of 3 dimensionsinto the vehicle,we could use a block of the
samematerial out of which we madethe z-dimensionalsheet,with
many electrodesembeddedin it to producevoltagedifferencesand
measurecurrents.But for 4 dimensionswe alreadyknow that we
haveto resort to 4-dimensionalnetworks, sincewe are not able to
make (or even imagine) 4-dimensionalblocks. In principle, this
could still measureshortest
doesnot make much of a difference.\U7e
42 | VEHICLE 8
vehicleshow off its built-in a priori conceptof higherdimensional
space.If the vehiclecould talk, we would ask it to roratein its mind
a 4-dimensionalcube,let us say90 degrees,aroundone of the axes.
There are such exercisesin human IQ tests,using z-dimensional
picturesof 3-dimensionaldice with three sidesshowing.The three
sidesare decoratedin differentways.The questionsareof this sort:
is cube A just another view of cube B, C, D, or E? Somehumans
havetrouble with 3-dimensionaldice, all with 4-dimensionalones.
But a vehicle endowed with a nerwork like the one in figure 15
might very well pass the IQ resr for 4-dimensionalcubesif the
questionwas posed in a languageit could understand.
I can hear myself talking to the philosophers.again.
The point I
am making is that orderly representationof spacein a vehicle is
more than iust convenienceof construction.It provides for easy
'\U7e
tests of reality.
have seenhow easyit is to knit nerworks that
will react to imagesmoving at certain speeds.If thesecan be taken
as imagesof obiectsin the world outside,the velocity ofthe movement of the imageswill stay benueencertain reasonablebounds, dictated by the physical laws gove-ing the movemenrof the obiects.
In particular, there won't be any movement of infinite velocity;
therewon't be any suddendisplacement.Continuity of movement,
no matter at what velocity, is a primary criterion for the physical
reality of an obiect. Also, the continuity and certainregularitiesof
the changeof shapeof a shadowindicatethat the shadowis castby
a solid obiect.This, too, could be fairly easilydetectedby a nenvork
with z-dimensional connectivity.And of courseidentity of shape
irrespectiveof movement (a strong clue for obiectskeepinga certain geometrical relation with a given vehicle)can also be detected
by such networks. we will take up this point againfrom a different
point of view in the next chapter.Here it was sufficientto show that
in our vehicle,just as in the physicsof relativity, the recognition,or
even the existence,of obiects is related to the dimensionalityof
space,internal and external.
VehicleI
Shapes
44 | V E HI CL E g
Figure 16
A detectorfor bilateralsymmetry.Thereis an array of elements
onto
which an image is projected. Elements symmetrically spaced with respect to
the midline enhance each other. There will be a strong activation of the
array for bilaterally symmetrical images.
o, Jtr
45 |
t1 t}tn^ -'ll
**l"ay
5a u''arP
,rt pl{: 6,
^'!+^'ihr,,[.'r,
1.o
Shapes
'rii{r;;
i,
moreactivethanwheno'ty or,fif themis
"?i";;Jd.tit
shape(witha verticaloi,
thatwhenthevehiclefacesa symmetrical
"
' -- v \ J
'*'ery'
)
46 | VEHICLE g
the vehicle toward the object or turning the vehicle awayfrom the
obiect.
Consider the first case: feedback that makes the vehicle turn
toward the obiect. An observermight say that our vehiclehas that
obiect on its mind or our vehiclepaysattentionto that object.Well,
what if the object is anothervehicle?What would the situationlook
like to that vehicle, and how should it react? Obviously the situation in which a vehicle seesanother headingdirectly toward it,
whether in an inquisitive, a friendly, or an aggressive
mood, is a
specialcaseand well worth specialattention.The detectorfor bilaterally symmetrical shapes,which we have iust described,proves
helpful here: we may connectit to the output in sucha way as to
trigger the mechanismsthat govern the appropriatereactionsto
"another vehicle facing me" or "another vehicle having me in
mind." (Perhapsone shouldreactivatethe beautifulterm "confrontation": fronts coming together, facing each other.) In fact, it is
clearthat bilaterally symmetricalconfigurationsin a natural world
containing only vehicles(and no other man-madeobiects,such as
churchesor monuments)would mostly signify just that: a parrner
in interaction with the observer.
is a relation bet'weenbilateral symmetryin sensory(especially vi
and the concept of "thour"
n of the
person
second
has been
e buildersof temples and churcheswho, by
symmetricalarchitecture,
evoke the presence of an a
ln conversatlon
. The sameprinciple
rved in
suchas orchids,adopt bilater
rical sh
in order to be acceptedas "partners" by insects
rs keyed to this type of symmetry.
I want you to note that somethingnew and very important has
crept into our discussionof a detectorwith bilateralsymmetry.\U7e
decidedto give our type 9 vehiclesa systemof connectionsbetween
correspondingpoints on their right and left sides.In order to ex-
47 | Shapes
plain how usefulsucha systemwould be,we had to invokenot only
the externalappearanceof other vehicles(which our vehiclemight
meet)but their behavior as well. Things are getting complicated:
we are no longer working on individualstaken by themselvesbut
on the membersof a community in which there are complicated
interactionsbetweenvehiclesof the sameor of different kinds.
Everyimprovementthat we invent for the latestbreedof vehicles
put in circulation will either force othersout of businessby a processof Darwinian selection(seeVehicle 5) or make otherschange
their behavior through learning (seeVehicle z). Of course,this
makes it difficult to foreseewhat will actually work out as an
"improvement." Sometimesthe net effectwill be contrary to what
we expect, due to unforeseenreactions of the environment. But
certain great inventions will survive all vicissitudesand will be
I suspectthat the detectorof bilatimmuneto all shrewd defenses.
eral symmtry,which providesinformation about "being in someone'sfocusof attentionr" belongsto this category.Evenin biology
with all its complicated interactionsbenveenspecies,the symmetry
detectorhas remainedof primary importance.An insectin search
of a sexual mate does not really care if it gets occasionallysidetracked by an orchid as long as its symmetry detector servesthe
right purposein the mafority of cases.
Other insects fall for different kinds of flowers, for those with
radial symmetry, like daisies.We can also construct radial symmetry detectorsfor our type g vehicles:thesedetectorsmight indicate singularities in the world, sources from which something
emanatesin all directions.A radial symmetrydetectorcould also be
basedon the fact that no movementis perceivedon approachinga
patternlike that of figure 17. The picture remainsidenticalto itself.
A fundamental category of form is periodicity. A repetitive pattern may signify many important situations.It may signifya collection of identical individuals.Then again, a periodic pattern left on
the ground may be the track of a vehiclemoving by somesort of
48 | VEHICLE s
Figure 17
A patternthat is invariantto changes
of scale.A vehicleapproaching
the
centerof the figurehasa constantvisualinput (providedwe makethe
6gurelargeenoughandthelinesinfinitelythin).Theabsence
of perceived
movement
may be usedasdiagnostic
for figureswith radialsymmetry.
49 | Shapes
ter as lateral inhibition. We have seen that such an arcay of
threshold devicesneglectscontinuous excitation and enhancescontrasts. It gives maximal output for patchesof excitation spaced
sufficiently far apart so that they won't disturb eachother by inhibition. For a periodic pattern, the spacingis determinedby the length
and strength of the inhibitory connections.If we test the lateral
inhibition devicewith striped patterns, we will notice that it gives
the sameoutput no mafferhow the stripesare orientedif the inhibition works in all directions.
Taken together,vehiclesof types 8 and 9 have provided much
new evidencefor our law of uphill analysisand downhill synthesis.
A problem that taxes the minds of psychologistswhen they deal
with real animalsor humans,that of inborn concepts,found many
solutions when we attacked it from the downhill, synthetic direction. We built very simple homogeneousnetworks and then discoveredthat they contain implicit definitions of such conceptsas
3-dimensionalspace, continuous movement, reality of objects,
multitude of obiects, and personal relation. More and more we
are losing our fear of philosophicalconcepts.
The exercisesin syntheticpsychologycontainedin this chapter
dealmostlywith visualinput. It is of courseeasyto imaginea priori
conceptsin other categoriesof input, suchasthe tactileor olfactory
inputs.It is quite elementaryto provide the vehiclewith detectorsof
aural periodicity. Th.y would detect various frequenciesin the
purely time-dependent(nonspatial)input derivedfrom one of the
vehicle'sears (microphones).The a prioris of frequency,the socalled resonators,have been basic to human auditory theory for a
long time.
Vehicle10
Getting ldeas
jr
I GettingIdeas
S z I V E H I C L Er o
ward the margin of the universe,it will alsohavenoricedthe neighborhood relationsamongpairs of theseobjects:the screwsare next
to the hill, the light next to rhe battery, and so forth. One day, after
enough excursions,the vehicle will suddenlyrealizethar all these
paired associations(A next to B, B next to C
Z next to A)
make senseif the whole situation is seenas a closedchain. The
vehiclenow has the idea of a finite boundeduniverse,with objects
A to Z marking the marginal closed line. Once this "image" or
"idea" is generatedout of individual itemsof knowledge,it is there
to stay. It may, in fact, be immediately recordedon the maps,
whose use we have discussed.If so, we will observethat forever
after the vehiclemovesaround much more expertly.
We must be careful,however,not to let the processof acquiring
new ideasinterferewith the detailedknowledgethat our vehiclehas
assiduouslycollectedand carefully storedin many associativeconnections during its lifetime. We know that this may happen in
humans who are overly dedicatedto the developmentof ideas.
Th.y tend to connectmany individual casesinto generalcategories
and then usethe categoriesas if they were things,losingthe potential for categorizingin other ways by rememberingeachinstance.
In the exampleof the discoveryof the margin of the universe,I
can seethis danger.The idea of a closedchainof obiectsmay be so
strong that it keepsthe imagesof theseobiectspermanentlyactive
in the vehicle's brain. The consequenceis that associationswill
developbetweeneveryobject on the margin-and everyother object on the margin. The serialorder that led ro the original ideawill
therebybe lost or at leastsubmergedin a systemof much stronger,
massiveassociations.The way out in our casewould be to let the
excitation circulate in the closed chain associations.This would
strengthenthe associationsrepresentingthe serialorder of the objects and would not allow cross-association
to develop.
Here are somemore examplesof ideasthat may arisein vehicles.
There are coins lying around on the floor in the universeof the
St I GettingIdeas
vehicles.Someof the coinsare decoratedwith a pictureof a human
headand others are decoratedwith a number. One of our vehicles
hasalreadylearnedro recognizeand to distinguishthe two typesof
coins.That is, there are distinctpatternsof activity, saynno different thresholddevicesbecomingactivewhen one or the other kind
of coin is seen by the eye of the vehicle.Now one of the coins
showinga human head is flipped around by the vehicle-and suddenly it showsthe number.This happensagain and againuntil, by
the learningprocessthat we havealreadyincorporatedin our vehicles,an associationis formed: "head, flipping, number." Of course,
the associationalsoworks the other way around. once the association is acquired,the vehicleknows that, after adding the action of
flipping to the sightingof "number," the picture of the headwill be
seen.It may also be reinforcedby the contraryexperience,
when the
flipping of coins showing the number revealsthe head.
\U7emay call the whole complex of head-flipping-numberand
number-flipping-headthe idea of a coin with nrro faces.It arisesin
the vehiclealthough the two facesof a single coin are never seen
together.The idea of a coin with two facescan ariseevenif there
are some coins around with human headson both sides,as long
as thesecoins escapethe vehicleduring the phaseof "getting the
idea."
Here's another example.Moving through a garden,a vehicle
finds out that flower number one of a row is a sourceof food,
flowers z to 7 are not (they are poisonous),flower number 8 is
againa sourceof food and so areflowers r j, zz, andso forth. After
a while it may happenthat in the brain of the vehicleonly one of 7
thresholddevices(connectedin a circular fashion)alwaysbecomes
activein temporal coincidencewith the finding of a sourceof food
in a flower. This is again "getting the idea": that particular
thresholddevicewill be associatedwith the food finding sysremwith the consequentadvantageof being able to predict sourcesof
food without havingto investmuchenergyin the processof sniffing
s4 | VEHICLEro
'We
must suppose,of course,that the time it takesfor one
around.
thresholddeviceto becomeactive after anotherhas beenactivated
is exactly the sameas the time it takesto get from one flower to the
next or, better still, that the advancementof activity by one step in
the ring of threshold devicesis triggeredby eachflower.
All of this is not complicatedin principle but boring to carry out
in detail. \U7erely on the processof Darwinian selectionthat, starting with the vehiclesof Wpe 6, has introduceda greatvariety of
different patternsof connectionsinto the vehicleswithout our even
noticing it (although*. do recognizethe vehicles'greatlyincreased
complexity of behavior).We canwell imaginethat the vehiclecould
get the idea "edible flower" evenif the only flowersthat wereedible
were those whose ordinal numberswere squareor whose ordinal
numberswere prime. Thereis, however,a complicationin the cases
of squaresand of prime numbers.If thesenumbersget too large,the
vehicle has to perform a long and intricate dance between one
flower and the next in order to find out whetherthe flower's number is square or prime, leaving marks on the earth and retracing
them accordingto complicatedrules.We haveseenthis before,at
the end of the chapteron Vehicle 5, which also had its limitations.
No such difficulry arisesif the vehiclehas to find out whether a
number is evenor odd, or whether a numberis a multiple of six or
of eleven,as long as the vehiclecan count to eleven.
In this chapter we were only interestedin the generalidea of
"getting ideas." Readerswho want to know exactlywhat kind of
to calculatenumbersthat
nerwork of thresholddevicesis necessary
are square,or prime or whatever,must read the textbooksof automata theory.
Vehicle
11
Rulesand Regularities
56 | VEHICLETT
things in the environment and establishestheir internal models
(called concepts).But this is not enough. These things move
around, bangagainsteachother, associateand dissociate,grow and
break.We havealtogethermissedthesedynamicaspectsup to now.
\Wewill introducethesedynamicsby improving on the systemof
Mnemotrix connectionsalready introducedin the type 7 vehicle.
You will rememberthat theseconnectionsbetweenelementsin the
vehicle'sbrain were of different strengthsand could be mademofe
effectivewhen the elementsthey connectedwere often activated
together.This turned out to be very convenient,becauseso manyof
the facts about the world that are interestingand important to us
(and to the vehicles)may be expressedasthings or eventsthat tend
to occur together.For this reasonit is unlikely that we will give up
the trick of associativelearningin any further developmentof more
refinedvehicles.
But we soon discoverthat there are important piecesof knowled{e about the world expressedin a different form: eventsthat do
not presentthemselvesat the sametime but in succession-pairsof
events,of which one is always the first and the other the second,
like lightning and thunder, swinginga hammerand hitting the nail,
or, in the world of vehicles,meetinga sourceof food and tastingthe
food. When we discovera pair of suchevents,we tend to think that
one is the causeof the other, whateverthat means.But this may
lead to wrong interpretations,for instancewhen both eventsare
producedby a third hidden event,only with different delays.Most
of the time, however,when two eventsregularly occur in succession, it is no accident.And it certainly is useful for a vehicleto
know what to expectwhen eventsoccurthat haveimportant,possibly dangerous,consequences.
\Uflecould useour old supply of Mnemotrix wire togetherwith a
little electronicsto incorporate into the vehicles'brains all those
delayedcoincidencesof eventswe have beendescribing.I7hat we
want to achieveis a connectionbetweenthe rwo internal represen-
57 | Rulesand Regularities
tatives of an event A and an event B such that, when the representative A is activated by the input, the representative B is activated by
the connection, but not vice versa. The connection would then
represent the fact that "B often follows A" or, if you wish, the
causal tie benveen A and B. This would force us to do a rather
complicated wiring for every such connection. In order not to burden our constructive imagination too much, w prefer to buy a
different sort of wire, called Ergotrix, which conducts in one direction only and has an increased conductance when it is interposed
berween elements that are active in successionwithin a brief time.
We must be careful, of course, to install the wire in the right direction, conducting from the element that tends to be active first to the
one that tends to be active second.
Once again we will see to it that all of this happens auromarically. Plenty of Ergotrix wire will be installed between as many
elements as possible so that whatever sequencesoccur can be recorded in the system. Of course there will be no lack of opportunity
for learning. tU7ithall the movement in the world around the vehicle, with all the natural laws operating, and with all the other
vehicles displaying fairly regular behavior on the basis of all the
tricks that we (or the processesof evolution) have built into them,
many sequencesof events will repeat themselves and they will be
worth learning.
You may ask why we did not use Ergotrix wire in the first place
(Vehicle 7) when we first gave our vehicles the capacity to learn,
starting with those complexes of properties that frequently occur
together becausethey belong to one "thing." We used the Mnemotrix wire, which is ideal for associations, because it couples elements in a symmetrical fashion; once coupled, each of the properties
can recall the other in quite the same way. For each Mnemotrix
connection we could have used two Ergotrix wires (one for each
direction) to obtain almost the same result. But there are two reasons to leave things as they are.
First of all, we don't want to go back in evolution and change
58 I VEHICLETT
things that have alreadyproved to be convenient,sincewe might
losesomeadvantagethat we havenot evenrealized.(Rememberthe
of uphill analysisand downhill synthesis:we run the risk of not
Z-law
any longer what we previouslyput together.)Sectunderstanding
probably
it
is
a good idea to keepthe nvo processes
ond,
conceptu4 t
ally separated-the associationsof elementaryproperties into
things or conceptson one hand and the sequencingof conceptson
the other hand, one controlled by the Mnemotrix, the other by the
Ergotrix system.The two kinds of learning produce two different
d
X kinds of knowledge,likegeographyand historyror systematiczool.-{ "W and animal behavior, referring to what kinds of things exist
*{:Jand
to how they develop and interact.
If we let our imaginationsgo and try to work out in detail what
kinds of things the Mnemotrix systemwill discoverin a real world,
and what kinds of dynamic laws will be incorporatedin the Ergotrix system,we soon discoverthat the two kinds of knowledgeare
perhapsrelatedmore than we had assumedinitially for reasonsof
conceptualconvenience.
First of all, it would seemthat the process
of abstracting things from the environment-concept formation at
the most elementarylevel-must occur prior to the processof discovering the dynamic properties of these things. For the laws of
successions
of eventsrefer to the developmentand to the combination of things rather than their elementaryproperties.This is familiar from our own human experience:listeningto a new language
we want to learn, we must first discoverindividual words, or roots
of words (somethinglike the morphemesin linguisticterminology),
before we can evenhope to discoverthe rulesthat governtheir use.
Also, in the developmentof a scienceit is often apparent how the
discovery and denomination of phenomenaprecedesthe definition
of the laws of their transformation.Chemistryhad to go through a
descriptive phasebefore the physics underlying the variety of subst'ancescould be understood. Zoology had to be taxonomic before
it was organizedby the theory of evolution.
V '
t .
59 | Rulesand Regularities
On the other hand, purely descriptiveclassificationis not only
boring, it is also potentially misleading.It may lead to the wrong
categorieswhen it is not guidedby at leastthe intuition of a theory
of the underlyingprocesses.
A centuryof microscopicanatomyhas
filled the librarieswith thousandsof beautifullyillustratedvolumes
that are now very rarely consultedbecausethe descriptivecategories of the old histology have beenlargely supersededby the new
6 o I V E H I C L Er r
describedthe conditionsfor the strengtheningof an Ergotrix wire.
These conditions are fulfilled when an element,say a threshold
device,at one end of the wire becomesactiveshortlybeforeanother
elementbecomesactiveat the other end. We havealso seenthat it
is mostly groupsof suchelements,stronglyinterconnectedand representing "things," that become active in succession.Now let's
introduce the rule that whenever the Ergotrix wires become
strengthened,the Mnemotrix wires within eachof thesegroupswill
also becomestrengthened.
Thus conceptsare establishedin the vehicleespeciallywhen they
appearin regularsequences.
How would this look to us?We would
notice, observingthe apparentlyerratic behaviorof a vehiclein its
world, that the vehicledisplaysparticularly well definedreactions
to eventsthat are known to haveconsequences.
Take, for example,
'Sfe
a vehicle approachingan obstacleat high speed. would not be
surprisedto seethe vehiclepromptly reactto its perceptionof the
danger of a collision. Similarly, Vehicle r r will quickly remember
which of its own behavior pafterns regularly and quickly elicit a
reaction from other vehicles.We observethat afteran initial learning period Vehicle r r will either produce thesebehaviorpatterns
frequently or pointedly avoid them. It will use them as signals.It
will also learn thosesignalsthat regularlyprecedecertainbehavior
patterns of other vehicles.After a while Vehicle r r will react to
thesepremonitory signalsiust as it reacted,beforethe learniqg,to
the behavior that regularly followed the signals.
But it would take prolongedobservationto noticethis particular
aspectof learningin the vehicles.As a matterof fact, we might not
have suspectedit if we had not introduced a piece of our own
philosophy into the construction of thesevehicles.As our brain
children becomemore efficient,we notice that the "law of uphill
analysisand downhill synthesis"becomesmore and more compelling. For the time being, take the messagein this form: sinceyou
were not satisfiedwith the first meagershowingof intelligencein
6r I Rulegand Regularities
our vehicles,we started adding a few more tricks, hoping that they
would convinceyou a little more. The first uick we tried was the
coding of the environment in thoseterms that yield a maximum of
correlations and logical structure, in other words, in the most
meaningful terms.
Vehicle12
Trains of Thought
I Trainsof Thought
64 | V E HI CL E tz
r __>
Figure 18
B is tlhebrain,whichreceives
input I andelaborates
an outputo. At the
sametimeit signalsthelevelof activityA in itsinteriorto a special
boxthat
calculates
appropriate
thresholds
@for the elements
in B.
for purposesof illustration it will sufficeif the thresholdcontrol
deviceworks just on the amount of aciviry in the brain.
The effect of this global negarivefeedbackon rhe activity of a
vehicle'sbrain is illustratedin figure 19, which showsthe number
of active elementsas a function of the numberof activeelementsa
'When
moment earlier.
the activiry is low, it will againbe low at the
next moment. (For very low excitation, there may even be a tendencyfor the activiry to die our, sincea minimum densiryof active
elementsin the brain is requiredto activatethe next setof elements,
but this is not shown in figure r9.) For very high levelsof excitation-that is, for a very large number of activeelements-we may
imagine that the thresholdsare immediatelyset so high that the
activity will drop to a very low level at the next moment. Intermediatelevelsof activiry will lead to maximum activity atthe next
moment (seethe middle part of the curvein Egurer9). Later on we
will come back to this curve, which has interesdngphilosophical
implications. First let us watch the operationof a brain that con-
Gl'b^)
!rr-,fl'jlu-h;r{,h^Yi
65 | Trains of Thought
Figure 19
giventhe
The function describingthe next number of activeelementsA;..,,1
presentnumberof activeelementsA;. It can be seenby iteration (follow the
linesstartingfrom the arrow neara) that the statesof a brain controlledby
sucha rule are quite unpredictable.
tains many learned associative connections while it is being controlled by the feedback of a threshold control device.
We have already noticed that the vehicle's brain has a tendency
to explode into fits of activity because of the abundance of reciprocal activation between its elements, a situation reminiscent of the
chain reaction in a block of uranium. But most of these explosions,
if everything works out the way we have planned, should take place
within limited groups of elements that are tied together by particu-
+c, t,
"An
R./ Lo,UJ.,
..J
4 ftrTe- a,,h,,tl
,,1-rrrl r.'"r,,{
u,$rrrf
vit
rov\""r+.
rections.Suchsetsof elementsariseas
"concepts" representingthings or eventsthat haveoften presented
themselvesin the environmenr.
Let one suchthing appearin the sensoryspaceof Vehiclerz. The
explosion of activity will happen in the correspondingset of
threshold devicesresponsiblefor that concept.This impliesan increaseof the number of active elementsin the brain, and the
thresholdcontrol devicewill immediatelyreact to it by raising all
the thresholds.A moment later many elementsthat werepreviously
activewill be silent. But the elementspertaining to the conceptin
questionare likely to stay active.This is becausethe strong reciprocal connectionswithin the set, once activated,guaranteea very
high level of excitation for eachelementof the set.This level is so
high that the activity of the elementsmay survivethe raisingof the
thresholds.Thus the first interestingeffectof our recentinnovation
is the focusingof individual concepts-of parternsthat havetheir
own internal consistency-at the expenseof backgroundactivity.
We greatlyappreciatethis effectin a well-functioninghumanbrain,
where it is often called the rocusrNc oF ATTENTToN.
But there is more. You rememberthat we haveinstallednot only
Mnemotrix wire for concept formation but also Ergotrix wire,
which representswithin the brain the relation of temporalsuccession, of consequence
or causality.Thus the elementsnow activein
the lone surviving concept after the automatic raising of the
thresholdsalso have someErgotrix wires attachedto them. These
Ergotrix wires lead to the elementsthat have often beenactivated
after the conceptin question,the consequences
of the activeconcePt'so to speak.Obviously, there will be more than one possible
next step for all but the most determinedsituations.
So we must ask ourselveshow the vehicle'sbrain finds the concePtthat follows the one it presentlyholds.The choice,it turns our,
is quite automatic.Among all the elementsactivatedby the presenr
concept through the Ergotrix wires, there will be some groups
67 | Trainsof Thought
strongly connectedby Mnemotrix wires becausethey again form
concepts.Thesegroupswill of courseignitewith particular alacrity
becausethe internal connectionswithin eachgroup will provide an
explosivekick to the activationfrom externalsources,that is, from
the active concept. Now you can see what will happen. The
threshold control, alarmed by all this growth of activity, will
quickly raise the thresholds,smotheringmost of the activity and
leavingonly the most resistantgroup of elementsactivated.As we
have already seen,this will be the group with the strongestreciprocal connections.In termsof conceptswe may put it this way: the
next concept, among all the concePtsthat are possible consequencesof the presentone, will be the most consistentor familiar
one-the one most strongly establishedby experience.
Note that with all these budding and growing explosionsthe
thresholdshave beenraisedabovethe level at which they were set
for the previousconcept.It is thereforevery likely that the previous
conceptwill be extinguished.So the systemwill not swing back into
its former condition but will end up with a different concept.This
embodiedin Ergotrix
new conceptwill haveits own consequences
wires. And thesewill againmaterializein a new concePtby way of
the sequenceof eventsthat we havejust described.The processwill
continue as long as you wish or as long as the chain of concepts
doesnot lead back to the conceptfrom which it started.
The upshot is something very much akin to thinking, to that
processso familiar to our introspection,where imagesappear in
successionaccording to rules reflecting the relations between the
things they stand for. This processgoeson in our minds when we
try to figure out the bestway to get from one point to another in
a familiar city by leaing our imagination produce successionsof
streetcorners(or other landmarks)whoserelationsof geographical
proximity we haveexperienced.It is alsoone of the tricks we useto
of possiblemovesin a game of chess,
determinethe consequences
of somestatementin a discussion.This chainor the consequences
hrf
68 | vnulcLEtz
ing of internal statesis exactly what we plannedto introduceinto
the brain of Vehicle rz to make its meditationslook more lifelike,
more like our own, not only in the time they take but also in the
unforeseenroutes they can follow.
There is an importanr properry that the brain of vehicle rz
shareswith the brains of our fellow men.consider againthe curve
of figure 19, which showsthe numberof activeelementsas a function of the number of active elementsa moment earlier.The exact
shapeof the curve is not very important, as long as it has a maxh,lh I imum and cuts the diagonal (A1: Ai*1). Startwith a certainvaluea
nrt
<4, > 1 on the abscissaand find the ordinate of the next value & on the
curve.Put that value b againon the abscissaand find c, and so on.
You will be surprisedto find that the succession
of valuesa, b, c
.
.
does
not
seem
to
follow
any
rules
is
and
in generalquite
r]
CWc
unpredictable.Now you will rememberthat figure 18 describesrhe
effectof thresholdcontrol on the activity of the brain of Vehiclerz.
r!(/emay takea, b, c .. . asthe numberof activeelements
in the brain
in successive
moments of time. If there are very few elements,the
succession
will by necessitybecomerepetitiousafter a short while.
But for a fairly largebrain the succession
will be truly unpredictable
to an observer,for any practicablestretchof time.
I hopeyou realizewhat this means.If you could observethe inner
workings of the vehicle'sbrain, say, by watchinglight'bulbsconnectedto the threshold devices,and theselight bulbs lit up every
time the correspondingelementbecameactive,you could not even
predict how many lights would light up in the next moment, let
alone what kind of pattern they would form. (For any givennumber there are of course many constellationswith that number of
activeelements!)At this point we'should againinvite our philosophers to comment.
I would claim that this is proof of rnnr \nLL in Vehiclerz. For I
know of only one way of denyingthe power of decisionro a creature-and that is to predict at any moment what it will do in the
69 | Trainsof Thought
future.A fully determinedbrain should be predictablewhen we are
informed about its mechanism.In the caseof Vehiclerz, we know
the mechanism,but all we can prove is that we will not be able to
foreseeits behavior.Thus it is not determined,at leastto a human
observer.
I know what the philosopherswill reply. Th.y will say that althoughthis may look like freewill, in fact it is not. \U7hatthey have
in mind when they use that term is the real power of decision,a
forceoutsideany mechanicalexplanation,an agentthat is actually
destroyedby the very attempt to put it into a physicalframe.
To which I answer:whoevermade animals and men may have
beensatisfied,like myself,a creatorof vehicles,with somethingthat
for all intentsand purposeslooks like freewill to anyonewho deals
with his creatures.This at least rules out the possibiliryof petty
exploitationof individualsby meansof observationand prediction
of their behavior. Furthermore,the individuals will themselvesbe
unableto predict quite what will happenin their brainsin the next
moment.No doubt this will add to their pride, and they will derive
from this the feeling that their actions are without causaldetermination.
Vehicle13
Foresight
And indeed-following up the last sentenceof the previous chapter-it may be said that the internal rumblingsof Vehicle
rz are at leastaimless,if not random, constrainedas they are only
by the rules of plausibility stored in the vehicle's memories
(Mnemotrix and Ergotrix) but not determinedby them.
I am sure that most of you will not believethat "aimlesssuccession of images" is an accuratedescriptionof what goeson in your
minds most of the time. You will not be impressedby our vehicles
aslong as thereis no evidenceof somepurposeguidingtheir behavior and some direction in their thinking. Theseare virtueswe are
pleasedto seein our children.Why shouldn't we try to modify our
brain children,the vehicles,in this direction?It won't be difficult in
principle, and it meansa lot to thosephilosopherswho like to think
that goal-directedbehavior is the one property that gives living
beingstheir very specialstatuswithin the physicaluniverse.
There are two aspectsof goal-directedbehavior we must consider.First, the goal lies in the future. For instance,the eatingof the
mouseis the goal determiningthe movementsof the cat now. We
have the specialcaseof an event defined for a later time having
earlier effects, quite contrary to the effects that we are used to
7r I Foresight
consideringin physics.Second,the goal is desirableby its very
definition.We cannottalk about goalswithout first getting straight
the conceptsof good and bad.
Let us take the first problem first, that of actingtoward the future
or in accordancewith an event in future time. This is obviously
nonsenseif we take it to mean an action that is now a consequence
of somethingthat will happenonly in the future. However, it is an
entirely different matter-and it doesmake sense-if we take it to
mean an action that is a consequence
of somethingwe expect to
happenin the future, sincethat expectationmay well be available
before the action is planned. There is no violation of the law of
causality in this. All we need is a mechanismto predict future
eventsfast enough.o that they will be known beforethey actually
happen.
There are of coursesafepredictions-and othersthat are not so
safe.We have no problem predicting the future of a rolling stone
onceit is on its way down the slopeof a hill. But it is not so certain
whether a dog will leaveits comfortablepillow when it is shown a
pieceof cake.Other motionsare practicallyunpredictable,like that
of a child playing in the middle of the street.Yet the principle of the
prediction is very similar in thesecases.We haveseenenoughrolling stonesand hungry dogs that the perceptionof one situation
immediatelybrings to mind its consequences.
Storedsequencesof
eventsare all we need for prediction, togetherwith a mechanism
forcing them to speedup in the reproductionwhen necessary,for
example,in dangeroussituations.Complicationsmay arise when
severaldifferent predictionsare approximatelyequally likely. In a
good prediction there must be the possibilityof predicting various
outcomes,given a certain situation, and of keeping the various
outcomesin mind in parallel. This is what we do when we drive
through a streetwhere children are at play.
Now we want to incorporatepredictioninto the vehiclesof type
r3. Clearly, the prerequisitesare all therein previoustypesof vehi-
72 | VEHICLETl
cles.We were careful to reproduceinsidethe vehicles'brainsmany
rules and regularitiesthat govern the world. This way we could
speakof the vehicles'brains as modelsof the world, as miniature
editionsof external,public space.Their brainswerepopulatedwith
patterns of activity that mimicked the activities of real obiects in
their environment.We noticed that thesebrains (as modelsof the
environment)really cameto life only when the dynamicaspectsof
the world were also represented,so that a given functionalstateof
the elementsof the brain would evolveinto the next stateaccording
to the samerules that make the world evolvefrom one momentto
the next. lUtledid this by using Ergotrix wire, which activatesthe
elementsof the brain in the sameorder asthe sequence
of eventsto
which they correspond.And we implicitly assumedthat the Ergotrix wires would be trained to reproducesequences
of activationat
the samepaceas the original occurrenceof the sequences
of events.
But this is a somewhatgratuitous assumption:the Ergotrix wires
could work faster,or slower,than the sequences
that are impressed
upon them. Let them reproducethe sequences
at a more rapid pace
and you will have a brain that works as a predictor (figurezo).
We want to take a closer look at what goes on in a vehicle
equippedwith such a predictor. Rememberthat the thresholddevicesin the brain are ugder the influenceof two kinds of input: first,
directly or indirectly (via interposedfilters) from the sensesand,
second,from one another. Only the latter kind of influenceis
mediatedby Mnemotrix or Ergotrix wires.Considera certainstate:
the vehiclein quiet contemplationof the world, the thresholdcontrol at rest, and the thresholdsset high enoughso rhat only a few
ideas stand out over the background. (Theseideas are of course
represented by groups of active threshold devices with their
Mnemotrix cross-connections,
)
The evolution of the vehicle'smental stare may be affectedin
three ways. First, meditation. Even if the brain is at equilibrium,
with the thresholdsfixed, it cannot be entirely at equilibrium be-
;2 to-1
prcdictor
t;3 $'2 tl
to td'l S2 q
Figure zo
A predictor with someauxiliary equipment. The flow of life is represented
by the film (or tape) being unwound from the reel marked future and
endingup on the reel markedpast. Only one momentof time, tro,is available
as input to the machine.The input is stored, however, for three units of
time on the endlesstape of a short-term memory. From there both the
presentinput and the contentof the short-termmemoryare relayedto the
predictor which computesthe future three units of time ahead.The predictor containsstatisticalinformation about the past embodiedin the Ergotrix
wires in its interior. The prediction for to * 3 is stored on another shortterm memoryuntil it is readyto be compared,by a specialcomparator,to
the real input t and to the input threeunits of time back.(Thisdepth in time
of the comparatoris desirablein order to assess
the dynamicpropertiesof
the predictor.) The comparator in turn emits signalsthat may modify the
predictor or switch it off (broken arrow).
T 4 | V E H I C L Er 3
causethe Mnemotrix connectionsbetweenthe activeelementswill
slowly grow in power, the longerthe ideais on. But this may not be
apparentfor a while, unlessnew elementsare recruitedto makethe
idea more ponderous,thereby upseaingthe equilibrium and making the brain go on a thinking tour, as we have alreadyseen(Vehicle rz).
Second,things may happen in the environment.The vehicle's
mental state will changeaccordingto new input from its sensors.
The transition from one stateto the next will be aided by the Ergotrix wires in the caseof a sequenceof eventsthat has occurred
to
before,but the Ergotrix connectionsare too weak by themselves
effect the transition without the help of the sensors'input.
Third, the sensorsmay signal a condition of the environmentthat
has alwaysevolvedin a certainway in the past. The Ergotrix connectionsin this casewill be very strong.And the next stateof the
vehicle'sbrain will be entirely determinedby them. As a resultthe
vehiclewill be blind to the real input that follows. Most of the time
this will not hurt the vehiclebecausethe sequenceof eventswill be
the sameas it has been in the past.
But occasionallythe rare event happensand the input clashes
with the internal prediction. This will result in a garbledcondition
that cannot developfurther in any coherentway. We want to avoid
betweenrealiry
this, especiallyin view of the fact that discrepancies
and expectationsare interesting and should be analyzedin detail.
Eventuallywe would like to provide the vehiclewith a devicethat is
turned on by iust these discrepanciesand amendsthe systemof
rulesusedfor the prediction,so that the vehiclewill know betterthe
next time it meetsthe samesituation.
First, we provide the vehicleswith two separaterepresenationsof
the environment,one in the predictor,the other in an equallylarge
ensembleof elementsthat receiveonly the fresh input from the
sensorsand do not elaborateon it. Thesenvo half brains are connected point to point to each other, so that the discrepanciesbe-
7S I Foresight
tween their states of activity can be detectedas easily as the
differencesbenrveen
two drawingsif you hold them one on top of
the other againstthe light. The technicalrealizationis easy.Saythe
two half brains are connectedby inhibitory connectionsbetween
correspondingpoints. Therewon'r be much activity if the two parternsof activationare exactlyequal,becauseof the reciprocalinhibition. But if one of the nvo representations
containssomeactivity
not presentin the other, this will stand out strongly.
\U7ewant our vehiclesto be imaginative,but mainly realistic.
That's why in the caseof conflictinginformation we want to take
the information from the realistic half brain more seriouslythan
that from the predictor.We may incorporatea rule: when in doubt,
believethe sensors.And we do this by introducing a mechanism
that simply turns off the predictor in casesof conflict.But we want
to go one stepfarther; we want to educatethe predictorto make it
more realistic.This is not as easyas it may sound.Rememberthat
the eventin the environmentthat causedthe predictor to make the
wrong prediction belongsto the past by the time the clashbetween
the two half brains revealsthe mistake.
Thus we want something like shoft-term memory (figures zo,
zt), a third representationof the environmentlaggingbehind the
other two, so that, if necessary,
the past is availableat any time a
few stepsback. such a mentalechois not difficult to incorporatein
the vehicle'sbrain. Just connectevery elementof the sensoryhalf
brain with another elementthat becomesactive one unit of time
after the first, and with yet another set of elementsthat becomes
activetwo units of time later, and you havean efficientshort-term
memory.
Now with a few additional piecesof equipment,we can greatly
improve the predictor by making it more flexibleand open to new
experiences.
We do not worry about occasionalwrong predictions,
especiallyif the mistakesare not fatal ones.Knowledgeis incorporated all the time in the Mnemotrix and Ergotrix connections,and
76 | veHtCLE t7
Figure zr
Learning by internal repetition of one-time evenrs.I is the input. D is the
Darwinian brain which emits judgmentson the desirability of the input and
sets two switches accordingly (through the broken arrows). On the left:
normal operation, with the Darwinian brain quiescent.The realistic brain
R feedsthe predictor P and also an open chain of nvo delayelementsd. On
the right: the Darwinian brain D signalsemotionalinput. The two switches
are thrown to the right, the predictor is no longer fed by the realisticbrain
but receivesthe contents of the nvo delay elementsd to which in turn it
givesouput. The information precedingthe emotion reverberatesthrough
the predictor and the delay elementsuntil the Darwinian brain calmsdown
again and setsthe switches to the normal position.
Z7 | Foresight
property"ggo/opeaceful"with the color greenand to reactindifferently to the sight of green,sincesooneror later an encounterwith
the greenmaverickis bound to take placeand the victim mustbe on
the defensive.It is betterthen to give specialweight to the rare but
decisiveexperienceand to consider green vehiclesas generally
bad.o
How is this done?We are talking about "good" and "bad" as if
theseconceptswereeasyto define.Of coursethey arenot, but there
is a way out of this difficulty. Rememberthe vehiclesof our earlier
models.Th.y were fairly simplemindedcomparedto the ones we
are now developing,but they were efficient. The type 6 vehicles,
which underwenta processof Darwinian selection,know one thing
for certain: the avoidanceof dangerand the searchfor advantage.
And they know this eventhoughno one (neitherthe vehicles'buildhasany idea of a definitionof good
ersnor the vehiclesthemselves)
or bad. The type 6 vehiclessimply move forward toward good
things and back away from dangerousthings. But this is all we
need.
Catch one of those Darwinian vehiclesof type 6, take away its
motors, and you have a detectorfor good and bad. The wire that
went to the forward motor signals"good" and the wire that went
to the backward motor signals"bad." So we can incorporatethe
brain of Vehicle 6 into the brain of Vehicle r 3 and therebyprovide
it with important, ancient,intuitive knowledge.
We can now put the piecestogether. Short-termmemory, two
stepsback in time for everythingthat happened,is alreadythere.
The predictor is there. A switch that momentarily turns off the
*I hope you rememberthat we are only talking about liale machines.It would be
wrong to cite the usefulnessof one-instancelearning in vehiclesas a iustification for
prejudice and superstition in human behavior. We do have vastly more complex
brains that enable us to make the diagnosis of good and bad independent of
superficialmarkers such as the shapeor the color of the casing.
78 | VEHICLEt3
predictor in the caseof a conflict berweenpredictionand reality is
also there.The Darwinian evaluatoris ready to signalparticularly
sinisteror ioyous events.The new trick: figure zr.
whenever the Darwinian evaluatorD signalsan unpleasantturn
in the real courseof events,or a very pleasantone, the predicting
half brain P is disconnectedfrom the input it normally receives
from the realistic(sensory)half brain, R. Insteadthe predictinghalf
brain receivesits input from the short-termmemorytwo stepsback.
Soit will go againthrough the rwo instanrsprecedingthe important
happening.At the sametime its output is connectedto the input of
the short-term memory. So it will receiveover and over again via
the short-term memory the successionof the two events,a and b,
until the Darwinian evaluatorD hascalmeddown and everythingis
switchedback to normal.
The net effect is that successionsof events leading to strongly
emotionalconsequences
are incorporatedfirmly in the MnemotrixErgotrix systemeven if they occur only rarely. The internal reverberationsetup by the Darwinian evaluatorarrificiallymakesup for
their low frequency and turns them into high-frequencyeventsin
the inner workings of the brain.
we may relax now and observevehicle 13 in action. Its power
of prediction is quite apparenr when it follows a moving object
around, say another vehicle carrying a sourceof attraction. When
the obiect temporarily disappearsbehind an obstacle,vehicle 13
will head toward the place where it is likely to show up again. we
also notice more peculiarproperties.For no obviousreason,Vehicle 13 seemsto avoid certainplacesand vehiclesin its environment,
and it seemsto have an irrational affection for someother places
and vehicles.If we watch it long enough,we may find out that there
are indeed reasonsfor these idiosyncrasies.The vehicle may associatea one-timeeventwith this or that placeor vehicle,and act
accordingly.Vehicle r3 remembersfacts much aswe do, individual
facts and eventsof its past experience.This rememberingis differ-
79 | Foresight
ent from the memorywe haveconsideredbefore,which consistedin
the molding of behavior accordingto the unchangingrules and
regularitiesof the environment, perceivedthrough the statisticsof
many individual events.The vehiclesof type 13 derivetheir experiencesfrom rare but important happenings.They will be quite different, one from the other, becauseeachvehicle builds up its own
characterbasedon the particular experiencesof its early life.
Vehicle14
As time goes on, we grow affectionatetoward the diversifiedcrowd of our vehicles,from the very simpleones to the
more complex modelsdisplayinginterestingsocialinteractionsand
sometimesquite inscrutablebehavior.we can play with them, we
may get to know them personally(and they may ger ro know us),
we can teasethem, testthem,teachthem tricks,and let themlove or
fight each other. We do not feel, however, that they show any
personaliry,not eventhe most complex onesof rype r3. It is difficult to say what we mean by that.
Perhapswe would acceptthem more readily as partnersif they
gave more convincing evidenceof their own desiresand projects.
we notice that our fellow men usually seemto be after something,
when they go about their businessor when we conversewith them.
Dealing with people is interestingbecauseof the challengetheir
continuousinternal schemingseemsto provide.The systemof desireswe suspectbehindtheir schemingmay be part of what we call
the personality. It may be the lack of just such projects that we
notice in our vehicles.we cannot help feelingthat they are driven
by necessityrather than drawn by goals-in spiteof all the efforts
8r I Egortstn
and Optimism
we put into them, in spite of specialmechanismsthat are apt to
abolishlowly forms of causaliry,and in spite of the predictor that
seemsto draw motives from a future stateof the world.
Once we have noticed this, we can of course,in a last creative
effort, endow a new kind of vehicle,our last, Vehicle 14, with a
certainamount of systematicegotism,with a touch of the pleasure
principle, in order to make it look more like our fellow humans.
We proceedas follows.
Rememberthat our more sophisticatedvehiclesalready have
built into them many componentsthat comein handy for this new
project. With the introduction of the Ergotrix wires (Vehiclero)
predictionbecameone of the vehicles'mentalhabits.In Vehicler3
the updatingof the predictor was greatlyimprovedby a mechanism
giving greatweight to rarc but imporrantevenrs.This was achieved
by incorporatinginto the brain of Vehicle13 another,more primitive, Darwinian brain that contributedall the ancientinformation
about good and bad things its ancestorshad accumulatedthrough
the generations.
Still earlier,we had noticed (Vehiclerz) that the succession
of
mental statesdictated by the Ergotrix connectionswas essentially
random and quite unpredictable(perhapseven unpredictableas a
matter of principle becauseof the peculiarmathematicalproperty
we associatewith the function of figure r9). The randomnessof the
decisionsmadeby Vehicle rz in part reflectedthe statisticalnature
of the knowledgeincorporatedin the Ergotrix connectionsand the
continuous updating of this knowledgeby an ongoing learning
process.It also dependedon the very nature of the processthat
makesthe brain swing from one stateof hctivity to rhe next during
alternateepisodesof raising and loweringthe thresholds,auromatically imposedby the mechanismof thresholdcontrol. IUe will now
givethis processan optimistic slantso that the pump of thoughtsin
the brain of the vehiclewill producea succession
of more and more
pleasurablemental images.We will convinceourselvesin the end
8z I VEHICLEr4
that such optimism not only leads to nice dreams but also has
obiectively favorable consequences.
Ve will assumethat most of the time the uncertaintyas to the
next state, given a certain state of activity, is not only an uncertainty for the observerbut an inherentuncertaintyin the sensethat
the predictor points toward (at least) two statesthat are equally
likely as a continuationof the presentstateof the brain (and therefore of the world). Sucha dilemma in previousvehiclesmight have
been decided by random elbment built into the brain (for ex"
ample, by Geiger counter making its decisionson the basis of
"
whether or not it was hit by a cosmicray within the last tenth of a
second).But from now on we will imposethe following rule for
Vehicle 14: when choosingamong severalequallylikely next brain
states,choosethe most pleasingone.
You have alreadyguessedhow we want to achievethis. We hold
the presentstatefor a short time (no problem,short-termmemory
is already there) while the predictor is allowed to go quickly
through its various predictions.At the sametime the built-in Darwinian evaluator is asked to evaluatethesepredictionsfor their
favorable or unfavorableaspects.It will in generalcome up with
this is done,the
different valuesfor the different predictions.'U7hen
predictor quickly goesonceagain through its predictionsand stops
This is
at the prediction with the highest scorefor pleasurableness.
then the next stateof the brain.
We don't needmore than that. We may put the vehiclesback on
the table and meditateabout their behavior.A superficialobserver,
or an impatient one, will not notice anything special. We, as
creatorsof vehiclesand experiencedobserversof their behavior,do
notice subtle changesin our latest perfectedbrain children. We
know their tastes:we have ample opportunity to seewhich sources
of stimuli, which situationsand which other vehiclesthey are attracted by and which they avoid. Their reactionsto thesethings in
the past were quite direct and easilyobservablewhen the obiectwas
A
j{;,
r\
hrV\,'''
t '
I
I
"
t
I
^t..
la^
r=r ///
) l5,iir
ll?
/,0
aa
aaQaaaao
a
a
a
a
a
ao
aa'
a
a
ot'a
a
a
a
aa :r.\
,
'"\'
a,'f
.:r.m
'//
.\,,d
| .:'
)b
v6
e
.l
tu,
,/6
,c
34t41434343434343{r
3131313rr3r3l313tll3l
-1
._o
T
I
Biologicd Notes
on the Vehicles
g7
| Crossed Connections
Figure zz
A simple explanation of crossed connections in the brain.'When a median
plane becomes defined in an animal with spherical symmetry, if each of the
elements is connected to each of the others, there are more fibers crossing
the median plane than fibers staying on one side. This little difference has
been invoked as the ancestor of the much more imposing crossed connections in vertebrates.
general rule.
Possiblythe bestknown explanationof the crossedrepresentation of the world in the brain is that providedby Ramony Cajal to
accountspecificallyfor the crossingof the optic neryes,which he
interpretedas a correctionof the image inversionthat occurs for
reasonsof geometrical
opticsin cameraeyes(r). His argumentis as
'follows.
Supposethe right and left halvesof the visual field are
projected,with optical inversion,onto the right and left retinae.If
thesetwo half imageswere projected by uncrossedfiber bundles
onto the right and left halvesof a commonreceivingsurface,there
would be a midline discontinuityin the mappingof the visual field
on this surface(figure4A). Ramon y Caial sawthe chiasmalcrossing as a simple meansof avoiding this discontinuity(figure 4B).
The other sensoryand motor systems,accordingto this theory,
adaptedsecondarilyto the crossedrepresentationof the world in
e8 | B T O LO GT C ALN OT E S ON T H E VE HTC LE S
Figure z3
connections,
staftingfromthe
of thecrossed
Ramony Caial'sexplanation
the continuityof the
inversionin the lenseye:the crossingreestablishes
on the brainin the awkarrowB, whichwould otherwisebe represented
ward fashionof A.
the brain. If visual imagesof objectsto the right are beingprocessed
in the left brain, it is economicalto let the motor commandsfor
actions dealingwith theseobfects(and presumablyexecutedwith
the right extremities)also arisein the left hemisphere.
Severalobjectionsto Ramon y Cajal's argumentmay be raised,
both on the basis of the reasoninginvolved and in the light of
experimentalresultssincehis time.
for correctionof optir. Crossingis sufficient,but not necessary,
cal inversion.For example,a rSoonadstof uncrossedfiber bundles,
or the equivalent internal crossingof fibers within each bundle,
would permit correction without inversion of the image (figure
z+:A,B). Similarly, recurvedand uncrossedbundlesproiectingonto
the posteriorpolesof the optic lobeswould also correctfor discontinuity (C).
an advantagein
z. The cogencyof Cajal's argumentpresupposes
a continuous unbroken representationof the visual field in the
99 | CrossedConnections
c
Figure z4
(A),or rwisting(B)of individual
crossing
A weakness
of Caial'sargument:
(C)woulddo thesametrick.
brain
of
the
to
the
back
detour
or a
bundles,
r oo I B I O LOGIC AL N O T E S ON T H E VE H TC LE S
Figure z5
Diagram of the eye and visual ganglia of the fly. Co is the cornea, a set of
lenses (in reality about 3,ooo). Re is the retina, with three light-sensitive
elements ("rhabdomeres," in reality 7 f.or each lens) arranged with their
tips in the focal planes of the lenses.Each lens projects an inverted image:
the complicated weave of the axons below the retina, Rax, compensatesfor
this and reconstructs the global picture on the first visual ganglion, La.
There are further inversions of the picture in the outer chiasm (Che) and
then again in the inner chiasm (Chi) benveen the second (Me) and the third
(Lba, Lbs) visual ganglia, but these are not readily explained on functional
grounds. Che is an example of an inversion that is not benveen the two
halves of the brain but happens separately within each half.
i/'
t/
I'r \
!'\'r
Figure z6
Explanationof the fiber pattern Rax berweenretina and brain. rlt) zt4,
projec,4j arepointsin visualsPaceseenby threeadjoininglenses.Their
the originalorder
tion is in the order tzr,4tL,543.The fibersreestablish
rzt45brate ancestors may well have been that between the nose and a set
of muscles used for locomotion, since in the water the business of
following chemical gradients is certainly important. The details,
however, are not clear. First of all, we don't know what kinds of
motors these primitive vertebrates used. If they were propelled
pair of fins, the case is analogous to that of our
primarily by
"
vehicles of type z, in the sense that the thrust produced by the
motor on one side of the animal (or vehicle) makes it turn toward
ro 4 | B I O LO G IC AL N O T E S O N T H E VE H IC LE S
oo
Figure z7
Crossing of visual (V) and tactile (T) input. Also the motor output
crossed (M). Only the olfactory input (O) is uncrossed.
the opposite side. The contrary is true for a fish, which relies mainly
on a bending of its body for locomotion. In this caserhe contraction
of the muscleson one side results in the animal turning to the same
side. Also, it is not clear whether it is advantageous for a fish to
turn toward the sources that activate its locomotion, a supposition I
used to explain the crossing between the olfactory input and the
motor output. Be this as it fr?y, this sort of explanation of the
crossedproiection of the world in the brain may have its merits. It
draws on a large body of observations on animal orientation and
locomotion under the influence of various chemical and physical
stimuli, with the older work well summarized in Fraenkel and
Gunn (8). one of the originators of this tradirion, JacquesLoeb, in
a succession of books propagated a mechanistic approach quite
similar to that of our vehicles (q). The outburst of zoological work
in this field was largely prompted by negative reactions to his ideas.
Besides Vehicles z and 3, the very simple, so to speak, one-
roj
I ObiectFixationin Flies
Figure z8
The turning tendencyinducedby movementin two oppositedirections
may not be of the samemagnitude(representedby the different lengthsand
directions of the arrows). This will produce a tendency to turn toward
wiggling objects on a stationary or void background. The differencebetween the forward and backward reactionmay vary accordingto position
in the eye (arrows).
ro 8 I B I O LO G T C A L N OT E S ON T H E VE H IC LE S
being completely determined by the initial velocity of the fly and by
the distribution of visual detail in the environmenr.
Another well-known instance of nonlinear input-output relations is apparent in the reactions of many animals to the sight of
other animals or moving objects. This dependsin a curious manner
on the size of the other animal (or object): small specimenselicit
prey-catching behavior, very large ones elicit flight, and intermediate-sizeobjects are examined in more detail. Somethingof this
sort has been shown even at the level of electrophysiological studies
of single neurons in the visual sysrem of the toad (r4).
McCulloch-Pitts neuronsdnd
redl neurons(Vehicle5)
Vehicle 5 is, of course,an embodimentof the old McCulloch and Pitts "Logical calculusof the ideasimmanentin nervous
activity" (rf ). This was one of the great boostersof modernbrain
science.Its experimentalbasisis in electrophysiological
studieson
the spinal cord.n The influenceof one input nerve ("posterior
root") of the spinalcord on one output nerve("anteriorroot") is
under certain conditions "monosynaptic": the fibersof the posterior root directly contactthe motoneuronsfrom which the fibersof
the anterior root originate.
*A glossary may be helpful for readers who are not trained in
the biological sciences. Neuron: a special kind of cell devoted to signal transmission in the nervous
system. Dendrites: usually ramified appendages of the neuron, which carry signals
toward the central part of the neuron. Axon: a single usually ramified appendage of
the neuron which carries signals away from the center of the neuron. Nerue: a
bundle of axons. Synapse: the place where the axon of one neuron transmits signals
to a dendrite (or cell body) of another. Motoneuron.' a neuron of the central nervous
system connected to a muscle. Sensory flet4ron: a neuron directly connected to, and
influenced by, a sense organ.
rrr
I McCulloch-Pins Neurons
Figure z9
Dependenceof the action potential on the intensity of stimulation. The
shapeof the rapid excursionof the potential E (abovethe dashedline) does
not changewith different intensitiesof stimulation, while the chargingtime
required to meet the threshold is shorter for higher intensity stimulation
(uppercurve).
II ?- I B I O LO G IC A L N O T E S O N T H E V EH IC LE S
of danger or in reaction-time experiments in a psychological laboratory, or in sports, must be governed by sequencesof very few
action potentials in the neurons of the motor system.tUfith neurons
producing action potentials at a frequency of the order of ro or roo
per second, the reaction to a stimulus that occurs in less than o.r
seconds must be triggered by the first, or by the first few action
potentials.
However, in many other situations, well studied by neurophysiologists, the signal within the brain corresponding to a sensory
stimulus is a burst of action potentials rather than a single action
potential. In such bursts, very commonly the frequency varies with
Figure 3o
The effect described in figure z9 applied to a small nerye net. In this figure
either A, B, or C alone, or two or three of them in combination, activate
the interneuron 11 which has a threshold equal to one unit of excitation.
But the more input elements that are active, the sooner interneuron 11will
produce its potential. Thus when A, B, C, and D are active at the same
time, coincidence of the output of 11 and 12 at E may not occur any more.
This is to show that threshold elements on a discrete time scalemay not be
an accurate description of neurons.
rrt
I Euolution
the intensityof the stimulus. Borrowing terms from computerengineering:there is an analogueprinciple involved in this which is
quite foreignto the digital operationof the McCulloch-Pitts nerve
net, or to the automation of automata theory. I(e are far from
understandingthe code or, what is more likely, the differentkinds
of codesthat are usedbet'weennervecellsin the brain. One point of
information theory, however, remainsvalid: all messages
can be
representedin theory by discretesignalson a limited number of
elements.This is the reasonwhy the vehicles'brains, made out of
thresholddevicesvery different from live neurons,may still display
somevery lifelike propertiesof information handling.
Most readerswill have recognizedthe vehicleleavingmarks on
the beachas a very elementaryversionof a Turing machine.For
thosewho are not familiar with this concept,I recommendeither
Turing's original articles or Minsky's book of. t967, or the very
friendly introduction in J. Sampson(zo).
The very last sentenceon Vehicle 5 must by necessityremain
cryptic,becausethe idea is not yet fully worked out.'What I refer to
is the increasein computing power of a brain that is endowedwith
the power of learning.No doubt this makesit possiblefor the brain
to write down its own record in the pattern of its interneuronal
connections,and then read it off again.But the way this is done is
very differentfrom the way a Turing machineworks with its tape,
printing, and readingheads.
Euolutioz (Vehicle6)
The gamewe are playing to generatethe vehiclesof type
6 hushesup most of the complexity of Darwinian evolution. My
aim was not to make propagandafor this theory. It is all too obviously correct for the people who are enchanted by its power
of demystification,while others will foreverinvent difficultiesand
rrj
I Memory
fi6
| B I O LO G IC AL N OT E S O N T H E VE HIC LE S
input
Figure 3r
Refinementof the projection of visual input onto the visual coftex by a
learning process.Fibers from both eyesreach the cortex in a rough topographicalorder, such that bundlesof fibersfrom correspondingplacesof
the two eyes are intermingled in the same compaftmentof the cortex.
Subsequentlyindividual cortical neuronspick fibersfrom both eyeswhich
are mostly active at the sametime (X) and make srrongconnectionswith
them (dots). Thus cortical neuronsbecomeconnectedto retinal elements
having exactly the samecoordinatesin the right and left eyes.
parts of the brain had the consequence that one of the two loci
yielded behavioral effects of stimulation previously associatedonly
with the other locus.
rzj
rz1 | Shapes
larger by at leasta factor of ro than the separationof their centers.
Their dendritic trees are fairly uniformly covered with synapses,
severalthousandsfor eachneuron,through which they receivetheir
input. (Theyare also connectedto eachother.) Thus eachresponds
to the activationof a cloud of synapses
centeredaround it, and the
belongingto neighboringneuronsoverlapgenercloudsof synapses
ously. Somedendritic treesof smallerneuronsare evenfully containedwithin the dendritic spreadof largerneurons(figure3z). We
:rt
Figure 3z
Fromcajal, r9rr. Golgipictureof the upperlayersof the humanvisual
coftex. only a small percentage of the total neuronal population is shown,
all of them of the pyramidal kind. The size of their dendritic ramification
varies a great deal. only part of the apical dendrites are shown for some
enormous neurons of the lower layers, the spread of which greatly exceeds
that of the other neurons in the picture (ascending dendrites marked c).
Figure 33
The neuronal zoom effect. Ve integrate visually over the hatching, which
admirably rendersthe smooth skin in the engravingabove.But shifting our
glanceto the animalsbelow, we are ready immediatelyto count the legsof
the centipedeor to describethe shapeof the book scorpion'sclaws, details
far finer than the spacingof the lines of the hatching.
rz7 | Shapes
variations in the density of quasi-parallelblack and white lines.
Individual lines are seenonly if artention is drawn especiallyto
them. on the contrary, if we glanceat the legs of the centipede
below, the segmentsof the bodiesof the three arthropods, or the
shapeof the clawsof the book scorpion,all their structural details
are immediatelyrecognizedas suchwith the full spatial resolution
our eyesafford. Note that the periodiciry of the centipedeis narrower than that of the hatching aboveand the small numbers next
to the zoologicalillustrationsare no largerthan the spacingof the
lines of the shading.
This observationimplies that we are able to switch rapidly from
one set of filters to another, making available to the formperceivingmechanismdifferent bandsof the space-frequency
spectrum. In termsof the neuronsin the cortex (figure 7z),itseemsthat
setsof smallerand largerneurons(by a factor of at least 5) can take
over in the coding of the visual input, dependingon which set of
neuronsprovidesthe picture that makesthe most senseto the brain.
Another categoryof visual perceptionis the continuity of lines
and of traiectories.Like clustering,it is implicitly embodiedin the
structureof nervenets as we seethem under a microscope;it also
providesgood evidencefor the usefulness
of internal maps. It certainly is not difficult to invent a network of "neurons" with connections benveenneighbors(figure 34) providingfacilitation suchthat
the input becomeseffectiveonly if one of the neighboringelements
has receivedinput a moment earlier.Sucha network would give a
much strongerresponsefor a patch of excitationmoving smoothly
over its surfacethan for disjoint patchesor discontihuousmovement. This is a common type of connectivity(for example,in the
system of axon collaterals in the cerebralcortex), although for
somereasonthe facilitatingconnectionsbetweenneighborsseemto
be lesseasilydetectedin the electrophysiological
experimenrsthan
l
)
.
t
1
n
the inhibitory ones.
-)
, vttupf
V
tl.w< :1 1n ,
r 2 8 | B T O L O G T C A LN O T E S O N T H E V E H I C L E S
Figure 34
A nerwork that responds to continuous trajectories. Neighboring elements
excite each other subliminally. They can be fully excited by the input (black
dots) only if a neighbor was excited a moment before. Thus only continuous trajectories (for example the one indicated by the stippling) are
perceived.
rzg I Shapes
a
.
1_
-\
(,
Figure 35
Contoursthat arenot presentin thepicturearereconstructed
throughan
activeprocess
of interpretation.
a: fromBrunswik,rg1'j; b: fromKennedy,
r974;c, d, ande: from Kanisza,
r974,all quotedin Metzger,r975.
in much the sameway for everyobserver,by someactiveprocess
that may have its roots partly in experienceand partly in inborn
mechanisms.We learn from this that it is somewhatartificial and
unnecessaryto draw a sharp line berweenperceptualand cognitive
(Sr).
processes
It would be surprisingif it turned out that the visualcategoryof
bilateral symmetry is not relatedto the symmetricalbuild of the
two halvesof the brain and to their point-to-point connectionsin
the simpleway I suggested
in Vehicle9. This very strongelementof
form, mirror symmetrywith respectto a vertical line situated in
front of the animal (jz), has an obvious counterpartin neuroanatomy: the commissuresconnectingsymmetricalpoints of the
rJr
I Shapes
Figure 36
Variouskindsof arrowheads.Their essence
is describedby the abstract
figureon the right.
r 32 | B I O LO G IC AL N O T E S O N T H E V EH IC LE S
Figure 37
Detection of discontinuitiesin the secondderivative.Five of thesecurves
are composedof arcsof circlesof different radii. Only one correspondsto
a single algebraic expressionof second degree.It is easily identified by
everyone(from Scheffers,
rgrr) (53).
It is tempting to think of Hubel and Wiesel's line segment detectors in the visual cortex as the elements of a differential analysis of
curves in visual space (l+). I7e are told that in any small region of
the cortex about half a millimeter across there are representatives
of all orientations in the set of line segment detectors, and there are
some that see white lines and others that see black lines or even
dividing lines bet'ween white and black. Since half a millimeter
cortex in the central part of the visual field corresponds to little
more than the unit of resolution of the visual system, we get the
impression that besides location and color, orientation is another
dimension in which the visual input is coded at the elementary level.
rtj
But we know very little about the possiblemechanismsof interaction betweenthe neighboringline segmentdetectorswe must Postulate in order to explain our proficiencyin the detectionnot only of
curvaturebut also of changesof curvature(figure,7), The feature
detectorsof Hubel and'Wieselup to now explain only first, not
secondand third derivatives.
{
I
()
2
ur
)
gro
e.
(J
F
!,
E
Ll
F
IJ
E
(,
a
234"b
DISTAilCE ALONGEIfCTR0OETRACXtmnrt
Figure 38
An a priori of music: the logarithmic plot of acoustic frequenciesin the
brain. Above (from Tunturi, ry62): projection of the frequency scale
(numbersstand for kilohertz) on the middle ectosylviangfrus (MES) and
again on the anterior ectosylvian gyrus (AES) of the dog cortex. Both
proiectionsare linear on a logarithmicscalebetweenabout zjo and 8,ooo
(16,ooo) herrz. Doubling of the frequencycorrespondsto equal distances
on the cortex. Only for very low frequenciesthis relation breaks down.
Below (from Evans, ry68)z a similar plot in the ventral cochlearnucleusof
the cat. Logarithm of frequencyversuspositionin the nucleusis linear (55).
45
learningand behavioron the one hand, and the introspectivepsychology of thinking (of which we are all specialists)on the other.
However, the distinction between two kinds of association,
which is introducedin Vehicle r r, is not only appealingon philosophical grounds but may find an interpretation in terms of
neuroanatomy.To make this plausible,I will provide an introduction to the cerebral cortex by quoting from some of my recent
papers(S6).
the numberof nerve
Accordingto somewhatdivergentestimates,
cells in both hemispheresof the cerebralcortex of man amountsto
about ro billion. The mafority of thesebelongto a type calledthe
pyramidal cell. It is characterized,amongother things, by an axon
rj6
('\/
Rt:)
r-.J
.+s
lL.
---D
I B I O L O G I C A LN O T E SO N T H E V E H I C L E S
r 3 8 | B I O LO G IC AL N O T E S O N T H E V EH IC LE S
special case I
ol fact.input
special casc 2
othcr sansory inpl
Figure 39
The skeleton cortex: pyramidal cells with long-range (A) and short-range
(B) connections. The olfactory input to the upper layers and other sensory
input to the middle layers are also shown (from BraitenberE, r97g).
49
Cell assemblies;embodiments
of ideas (Vehicles7, ro)
Vehiclesro ro 14 operarewith ideasthat standfor things
or situationsin their environment,and the ideasare represented
in
r4r I CellAssemblies
the brains of the vehiclesby groups of active elementsthat are
somehow tied together by reciprocal activating connections.\U(e
may ask: why are things not representedby single elementsor,
quite abstractly,by patterns of activity which are not constrained
by the condition that the active elemenrsexcite each other? The
answer is of course: becausethe vehiclesare caricaturesof real
brains and at presentit is again fashionableto think of groupsof
connectedneurons,so-calledcell assemblies
(64), as the carriersof
individual itemsof meaningor, if we wish, asthe morphemesin the
languageof the brain.
It is impoftant to realizethat there is no logical reasonfor this
interpretation. Supposethe point is internal representationof
thingsin sucha way that any largeenoughsubsetof the detailsthat
characterizea thing will be sufficientto evokethe thing in its entirety. This is undoubtedly a good principle, since it makes for
economicaluseof channelcapacityunder the assumptionthat the
existenceof things is the main redundancyin the world, with
"things" standing for bundles of details each of which, when it
presentsitself, raisesthe probabiliry for the rest of them to present
themselvesas well.* Under thesecircumstances
a Hebbian assembly of all the neuronswhich individually representthe detailsof a
thing is indeeda good codewordfor that thing, sinceit representsin
its internalexcitatoryconnectionsthe conditionalprobabilitiesthat
characterizethe thing itself. But a singleneuronthat receivesexcitatory synapsesfrom a set of afferenrsrepresentingthe detailsof a
thing is also a good codeword, vry much for the samereasons,if
its thresholdis set appropriatelysomewhereberweenthe value for
"all afferentsactive" and that for "one afferentactive." It, too, will
*If you want to grasp the meaning of terms such as
channel capacity, redundancy,
and conditional probability, you must read more information theory than I can pack
into a footnote. There are many introductory
Shannon and l7eaver's Mathematical
Theory of Communication
(tg+g).
r43 | ThresholdControl
neuronsas the rest of the brain. Theseought to be richly connected,
with a high divergenceof signalsfrom eachneuron to as many as
possibleother neuronsin the cortex.We havealreadyseenthat the
pyramidal cells do their best in this respect.This divergence(and
correspondingconvergence)is necessaryin order to provide as
much freedomaspossiblefor the choiceof partnersin the development of cell assemblies.
Finally, most synapsesought to be excitatory, sincecell assemblies
areheld togetherby excitatorysynapses.
About three out of four synapsesin the cortex are of type I, presumablyexcitatory (6g1.
)hot^rnvf
Relercnces
r.
L. Gaze,R. M. 1958. The representationof the retina on the optic lobe of the
ftog. Quart. ]. Exp. Physiol. 4rt zo9-Lr4.
t. Braitenberg,Y.1967. Patternsof proiectionin the visual systemof the fly. I.
Retina-Laminaproiections.E*p. Brain Res.3: z7r-298.
4. Autrum, H. J., and I. Wiedmann.1962.Experimentson the opticsof the insect
eye (appositionaleye). (Versucheriber den Strahlengangim Insektenauge(Appositonsauge).)
Z. Naturforsch. ry: 48o-482.
Kirschfeld, K. 1967. The proiection of the visual environment on the array of
rhabdomeresin the compound eye of Musca. (Die Projektion der optischen
Umwelt auf das Raster der Rhabdomereim Komplexaugevon Musca.) Erp.
Bruin Res.3: 248-27o.
j.
I
Paris tzz-r24.
6. Horridge,G. A., and I. A. Meinertzhagen.ry7o.The accuracyof the patterns
of connexionsof the first- and second-orderneurons of the visual systemof
Calliphora.Proc. Roy. Soc.B. ry5: 69-82.
1 46 | RE F E RE N C ES
Progressin Brain Res.
7 . BraitenbergrY. ry65. Taxis, kinesisand decussation.
r7i 2ro-222,.
Braitenberg, V. 1968. On Chiasms. lnz Neural networks, edited by E. t(.
Caianiello. Berlin, Heidelberg, New York: Springer-Verl"g,pp. 34-42.
Fraenkel, G. S., and D. L. Gunn. ry6r. The orientation of anitnals. Kineses,
taxes and cornpdssreactions. New York: Dover PublicationsInc.
[.ocb, J. rt9o. Hcliouopism in animals and its analogy with plant heliotropism. (Der Heliotropismus der Tiere und seine Ubereinstimmungmit dem
Heliotropismus der Planzen.) Wiirzburg. p. rr8. Quoted in Fraenkel and
ro.
Gunn, 196r.
Roessler,O. E. r98r. An artificial cognitive-plus-motivationalsystem.Progr.
Theor. Biol. 6z t47-t6o.
Koshland, D. E. r98o. Bacterial chemotaxisas model behauioralsystez. New
York: Raven Press.
II.
Fermi, G., and V. Reichardt, ry67. Optomotor reactionsin the fly, Musca
domestica.Dependenceof the reaction on the spacefrequency,th velocity and
the luminance of moving periodic patterns. (Optomotorische Reaktion der
Fliege Musca domestica.Abhengigkeit von der Reaktion der Wellenlinge, der
Geschwindigkeit,dem Kontrast und der mideren l-euchtdichtebewegterpe-
tL.
rr.
r09-20t.
Ewet, J. P. r98o. Neuroetbology. An innoduction to the nanrophysiological
fundamenuls of behauiozr. Berlin, Heidelberg,New York: Springer-Verlag.
r 5 . McCulloch, W.S., and V. H. Pins. r94j. Alogical calculusof ideasimmanent
in nerryousactivity. Bull. Math. Biophys.Szrr1-r1r.
1 6 . Creed, R. S., D. Denny-Brgwn, J. Eccles,E.G. T. Liddell, and C. S. Sherrington. tgtz. Rfiex actiuity of the spinal cord. Oxford: Clarendon Press.
1 7 . Eccles,l. C. 1964. The physiology of syndpses.Berlin, G<ittingen,Heidelberg,
I4.
r47
| References
r48
REFERENCES
,r.
,2.
3r.
r49
| References
t4.
Physiol.ryT 377-407.
B., and W. Reichardt.r955. Systemtheoreticanalysisof temporal
16, Hassenstein,
factors, of sequenceand sign in the perceptionof movement by the beetle,
Analyse der Zeit, Reihenfolgenund VorChlorophanus. (Systemtheoretische
des Riisselkiifers Chlorozeichenauswertungbei der Bewegungsperzeption
phanus.)Z. Naturforcch. rtb: Srt-1z'4.
Hassenstein,B. 1958. Perceptionof movementof figural and irregular patterns. (Uber die lgahrnehmung der Bewegungvon Figuren und unregelmi8igen Helligkeitsmustern.)Zeitschrift f. uergl. Physiol. 4o: 556-592.
Reichardt,W., and D. Variu. 1959.Transferfunctionsin the visualPerception
im Auswertesystemftir das Beweof movement.(Ubertragungseigenschaften
gungssehen.)
Z.f. Naturforschung.4b $o): 674-689.
37. Fermi, G., and W. Reichardt, ry63. Optomotor reactionsin the fly Musca
domestica.Dependenceon wave length, velocity, contrast and averageluminance of moving periodic patterns. (Optomotorische Reaktion der Fliege
Musca domestica.Abhengigkeit der Reaktion von der Wellenliinge, der Geschwindigkeit, dem Kontrast und der mittleren leuchtdichte bewegterperiodischerMuster.) Kybemetik zz 15-28.
Reichardt,V., and T. Poggio.1975.Visualcontrol of orientationbehaviourin
the fly. Part I. A quantitativeanalysis.Quart. Reu.Biophys. 3; 1rr-375.
Poggio,T., and W. Reichardt.ry76. Visualcontrolof orientationbehaviourin
the fly. Part II. Towards the underlying neural interaction. Quart. Reu. Biophysics9, jt 177-4j8.
38. BraitcnbcrgrY. r9Tj. On the texture of btains.New York, Heidelberg,Berlin:
Springer-Verlag.
j9. Mach, quotedin Radiff, F. 1965.Mach bands.San Francisco:Holden Day.
40. Bekesy,G. v. r95o. Experiments in hearing.New York: McGraw Hill.
4r. Hartline, H. K., and F. Ratliff. r957. Inhibitoryinteractionof receptorunits in
the eyeof Limulus.J. Gen. Physiol.4o: 757-367.
42. Reichardt,V. 196r. On the optical resolutionin Limulus. (Uber das optische
von "Limulus.") Kybemetik r 59-69.
Aufl<isungsvermcigen
Variu, D. 1965. On the theory of lateralinhibition. lnz Cyberneticsof neural
processes,
editedby E. R. Caianiello.Rome:C.N.R.
r5o I REFERENCES
Szentagothai,J. ry67. ln: Eccles, Ito and Szentagothai:The cerebellumas a
neuronal machine.Berlin, Heidelberg, New York: Springer.
pulses.
44. Beurle,R. L. r956. Propertiesof a massof cellscapableof regenerating
Proc. R. Soc.Lond. Ser.B. 24o: jj.
4j.
4j.
:J
I
+i
Wilson, H. R., and J. D. Cowan. r97j. A mathematicaltheory of the functional dynamicsof cortical and thalamic nervoustissue.Kybernetik 13 3j,
Hubel, D.H., and T. N. lViesel.rgTz.Laminarand columnardistributionof
geniculo-corticalfibers in the macaquemonkey.J. Comp. Neurol. l.46: 4zr-
450.
Probleme.Berlin: Springer.
46. Krihler, W. rg33. Psychologische
ITertheime4M. t923. Experimental contributionsto the theory of form. (Untersuchungenzur Lehre von der Gestalt.) Psychol.Forsch.4.
Petermann,B. 1929. Die Vertheimer-Koffka-K<ihlerscheGestalttheorieund
das Gestalproblem. l*ipzig: Barth.
For recent developmentssee:
Gibson,J.J. rgSo. Thefirceptionof the visualworld.Boston:Houghton Mifflin
Co.
tr\l
f"
\)
jr.
rjr
I References
53. Scheffers,G. r9rr. Lehrbuch der Mathenatik, z,ndedition. Leipzig: Veit and
Co.
54. Hubel,D. H., and T. N. Wiesel.1959.Receptivefieldsof singleneuronesin the
cat's striatecortex.J. Physiol.(Lond.) r48: 574-59r.
Hubel, D. H., and T. N. Viesel. ry77. Fundonal architecnrreof macaque
monkeyvisual cortex. FerrierLecture.Proc. R. Soc.Lond. B. r98: r-59.
5j.
rsz I REFERENCES
Evans,E. F. 1958.Uppe.rand lower levelsof the auditorysystem:A contrastof
structure and function.ln: Neural netutorhs,editedby E. R. Caianiello.Berlin,
Heidelberg, New York: Springer.
Aeftsen, A. M. H.J., and P. I. M. Johannesma.r98r. The spectrotemporal
receptive field. A functional characteristicof auditory neurons.Biol. Cybem.
42| r33-r4t.
62. Hubel, D. H., and T. N. Wiesel. ry65. Binocularinteraction.Seenote 27.
Wiesel,T. N., and D. H. Hubel. 1965.Comparisonof the effects.Seenote 27.
Blakemore,C., and G. F. Cooper. r97r. Modificationof the visual cortex by
experience.Brain Res. 3t: 766.
Hirsch, H. V. 8., and D. N. Spinelli. r97r. Modificationof the distributionof
receptive 6eld orientation in cats by selectivevisual exposureduring development. Exp. Brain Res. 13: r-43.
68. Braitenberg,Y. 1978.Cell assemblies.
Seenote 56.
69. Wolff, J. R. ry76. Quantitative analysisof topographyand developmentof
synapsesin the visual coftex. Exp. Brain Res.Suppl.r 259-263.
Uchizono, K. ry66. Characteristicsof excitatory and inhibitory synapsesin the
central nenroussystemof the cat. Nature zo7: 642.
Braitenberg,Y. 1978. Cell assemblies.Seenote 55.
Braitenberg,Y. 1977,On the texture of brains.Neuroandtomyfor the cybernetically tninded. Berlin, Heidelberg, New York: Springer.
7 r . Palm, G. 1982. Neural assemblies.An alternatiueapproachto artificial intelligence. Berlin, Heidelberg,New York: Springer-Verlag.
Ct*'t6i
\1q "'
Cu,l4
hvclR r