A Theory of Memory-Minsky

8/12/2019 A Theory of Memory-Minsky

1/24

M SS CHUSETTS INSTITUTE OF TECHNOLOGY. . , ., _ i . _ . , _ , . ' " _ . , . : ' __ , - jRTIFICI L lNTELLIGENCE L BOR TOlW

A ~ I . Memo No. 516 June, 1979

K-LINES: A THEORY OF MEMORY

Marvin Minsky

ABSTRAct. Most theories of memory suggest that when we learn rJ t l e t r i o r i ~ e something, some representation of that something is c o n s ~ r u c t e d tstored and later retrieved. This raises questions like:

How is information represented?How is it stored 1ltow is i t retrieved?Then, how is it used?

This paper tries to deal with all these at once. When. you gpt an idea andwant to remember it, you create a K-line for it. When later activated,the K-line induces a partial mental state resembling the one that created itA partial mental state is a subset of those mental i l I ; J n c h ~ s o}Jcrating at onemo#tent. This view leads to many ideas about the development, structureand physiology of Memory, and about how to implement f rame-l ikerepresentations in a distributed processor.

Ack.noWledgements: This report describes research done at. the ArtificialI n t ~ n i g e n c ~ Laboratoryof the Massachusetts Institute of Technology, Supportfor the laboratory's artificial intelligence research is provided in part by the .OfficEiofNaval Research under Office of Naval Research contract NOOO14-79-C-0260,

~ M A S S A , f W S l : n s 1 '5 . I. . . rm,IE OF. TE mSoWGY ~ 9 1 9


2/24


3/24

K LINES: THEOllY O MEMORYMarvin Minsky

M. I T.

Most theories of memory suggest that when you learn or tnemorizesomething, t representation of that something is constructed, stor>d and laterretrieved. This leads to questions like:

How is the information represented?How is it stored?How is i t retrieved1How is it used?

eW situations are never exactly the same as old. So i the information inaf io ld memory is to be useful, i t must somehow be generalized or

a b f ; t r a c h ~ d This leads us also to ask:HoW-are the abstractions made?When --before or after storage?How are they later instantiated?

We try to deal with all these at once, by developing tho thesis: thf u n ~ i o n of a memory is to re-create a state of mind. Then each memoryititlstembody information that can later serve to r e - a s s ~ m b l e the mechanismsthat were active when i t was formed -- to recreate a memorable brainevent. (See Note 1.) Our scheme is basically simple:

When you get an idea , or solve a problem , or have a memorableexperience , you create something we shall cali a K line for it.This K-line g e t ~ c o n n ~ ~ t e d to those mentalagend)s that were recentlyactive .:. Le. which were involved in the memorable mentaleveni.When the K-line is later a c t i v a t ~ d , t reactivates t h o s ~ mentalagenCies; creating a Ijpartlal mental state resembling the original.


4/24

3

TOfilaite this concrete, we must explain:What are "mental agencies"?lIow do K-lines interact with them?What is a "partial mental state"?How does this relate to "meaning"?

ISPOSITIONS vs PROPOSITIONS

In t,1s modern era of ';infonnatioil processing psychology'" i t may seemquaint to talk of mental states, The concept of represeniation ofknowledge seems lucid enough whf:m talking a1Jout memories of sentenceS,rttimbets . or even faces, for one can imagine how to fornp.11Clte these hiterrds of propOSitions, frames, or semantic Mtworks. But i t . is much harderto do this for feelings, insights and understandings, with all the attitUdes,dispstfions,and "ways of seeIng things" that go with them. See Note 2.tradition.dly, such issues are put aside, with the eXCuse that we should

u n d e t s t ~ n d simpler things first. But what i feelings and viewpoints arethe simpler things - . the elements of which the others are cOmPosed?. Then, ' I a 5 s e r t , w ~ should deal with dispositions directly, using a sttucthralapproacl1 thatPGrtrays memory as re setting the states of parts of the

n ~ f v o u s system.WewiU view a lllemory as something that predisposes the' mind to deal

.with. a I 1 ~ W situatio.ii. in an old, n ~ t n e t n b e r e d , way. This is why 1 put. d i s P ( ) ~ U l o n s ahead of p r ~ p o s i t i o n s ' ; . First we propose sotne dispositionr e p r e ~ e n i l n g structures. Then we try to showlhat these can evolve intothetItorefamiliar kinds of cognitive constructs we knoW as adults, . One" ,should, not assume that human memo.ryhas the same:untform, invarlan.t

c h a t ~ i f c t - e r throughout' deVelopment, not attribute to titrants abilities thatdevelop only later. Our' first model might serve for an infantile,dispos itiona l memory. Later we try to see how i t might evolve into a moreadu1t.system.

MEN1:ALSTATES nd the SOCIEtY o/MiNDorie'6u14say b.utUttle about "me;ntal states". if'one imagined the Mind {o,be

. asirl.gle, unitary thhig. But i we envision' a mind (or brain) as composed.o t m ~ l l y p a r t i a l l y autonomous "agents" -- a "Society" of smflllur minds

' t h ~ n We.tan interpret itmental state" and "partial mental state" in t.erms ofSUbSEftsof the states of the parts of the mind. To develop this idea, we willjm.aglnef lrs t that this Mental Soctetywork.s much like any human. atnbdstrative otganizaiion. .

,


5/24

Marvin lV1insky 4 ....;K;.:;. _T:;..:h:.:..;8ory of Memory

on the largest scale are gross Divisions that specialize in such areas assensory processing, language, long-range planning, and so forth.Within each Division are multitudes of subspecialists -- call themagents -- that embody smaller elements of an individual's knowledge,skills, and methods.No single one of these little agents knows very much by itself, buteach, recognizes certain configurations of a few associates and respondsby altering its s t a t ~ .In the simplest version of this, each agent has just two states, activeand qUiet. A total melltal state is just a selection of which agents areactive. A partial mental state is a partial such specification: i t f ixes. hesta tes of just some of the agents

t is easiest to think of partial states that constrain only agents within asingle Division. Thus one could think of a partial state that s p ~ c i f 1 e s some"visual imagery" without saying anything about agents o u t . " i d ~ the visualdivision. In this paper our main concern will be with e v ( ~ n smaller partialstates, that constrain only some agents in one DiVision.Notethat . the concept of partial state allows us to speak of entertainingsevefaipartlal states at once -- to the extent they ate compatible -- that is,they do not assign different states to the saine individual aomts. Even ithey conflict, the concept may still have some mr'aning, i thl" conflicts canbe settled within the Society. This .could he important, hecause local. JileChanisms for resolving d i f f ~ r e n c e s could be Ole antecE'dents of what weJulOW later as re soning -- useful ways to combine different fragments of. k r t o w l e d g ~In.the hext few sections we postulate that certain units -- the K-nodes andK.-lines a r e the elements of memory. When activated, Pilch such uni timposes. a specific partial state upon the Society. Such pUccts are notalways easy to describe, for we are most fluent at talking of arrangementsot sights and sounds -- or .motor patterns; these are "concrete" matters.Much more elusive se.em our recollections of attitudes, points of view, andfMlings. This does not mean that concrete recollection is fundamentallysimpler I t may only reflect the enormous competence of the logical andlingUistic parts of the adult mental society to communicate allout concretematters. That illusion of simpJicity can fool us, as theorists, into tryingfirst to solve the hardest problems. (See Note 3.

~ ~ ~ ~ '' '_........_- IIirI,...


6/24

The novice remembers being at a concert, and something of how t affectedhini.. The amateur remembers nlOre of what it sounded l ik( .Only theprofessional remembers much of the music itself, timbres, tones and textures.So , th e most concrete recollection may require the most refined expertise.Thus, while our theory might appear to put last things first, I maintain thatattitudes do really precede propositions. feelings come before facts. Thisseem.s strange only because we cannot remember what we knew in infancy.

MEMORIES nd PARTIAL BRA1N STArESOld ansWers never perfectly suit new questions, except in the most formal,logical circumstances. To deal with this, theorists have tried various ideas:

II Encode memories in abstract form.Search all memory for the nearest match .Use prototypes with detachable defaults.Remember methods , not answers.

Our theory resembles the latter. We propose remembering not the stimulusitself but part of the state of mind t caused. o we shall translate

into a method for solving a kind of problemonce solved a similar problem. f 1 can get into tllat

old state, I could probably handlc this Olte the saine lvay.To carry .out the translation we must sketch some of the atchitecture inwhich our Agents are ~ m b e d d e d (See Note 4.) We envision a braincontaining a great lattice of Agents , each connected to just a few others.biot t f modl31. we shall suppose that each agent's inputs come .either frombelow r from the side, and its outputs go 'Upwards or sideways. Thus

l n f o r i n ~ i o n can move only upwards, on the whole. (See Notc 5.) This iswha.t. one might imagine for the lower levels of a visual system, withsimple feature- and texture-detectors at the bottom, then edge- and regionsensing agents, then identifiers of larger structures.


7/24

M..... ... ...- i 'n__..M.:;..;i;.;;;.n:...;.s.;;.;:kx.y__________ .:::6_______K-Theor,} . of Memory:

Given these constraints, i f one looks down from the viewpoint of a givenagent P, one will see other agents arranged roughly in a hierarchicalPyramid:

//

// /

/ 1\

/ /

/ \/ \/ \ P PYRAMJO

\P \\ \

/ \ \\ \

/ 1\ \\ \ \

/ \ / \ \\ / \ \

\ / \ \\ 1\ \ \

\ / \ / \ \ \\ \ / \ \ \

\1 emphasize that the network as a whole need not be pyramidal; the Ppyramid we speak of is an illusion of an agent's perspective.

CROSS EXCLUSION and PERSISTENCE.

In o t concept of the SOCiety of Mind, most agents are grouped in smallcross-exclusion arrangements. Each sends inhibiting conn(1ctions to the

others 'in its group, so that i t is hard for more than one to bf,l active'; at at ime . This kind of sub-structure, familiar in physiology, makes i tparticularly easy to re-set the state of a system; one need only force to onone Agent in each cross-exclusion group. Then that agent will inhibit itsassociates -- reducing their inhibiting effect on itself.The result: networks composed of cross-exclusion systems have a kind ofbuilt-in short-term memory. Once such a system is forced into a partialstate, even for a moment, then that state will tend to persist -- except for

t o s ~ agents under strong external pressure to change. j Accordingly, such asystem tends to have internal persistences. To an ou'tside ollserver. 'thesewill appear as dispositions -- distinctive styles of behavior. To make a

l r g ~ change in such a disposition, one has to change many of the agents'states. Small changes will only slightly perturb the overall disposition.


8/24

Marvin Minsky 7

Finally. we suppose that agents at the lowest Ipvpls tend to change statesmost frequently, in response to signals a s c e n d i n ~ from the outsirle or fromother P-nets. In the scheme described below, tlle states of intNInediate levelagents will have the most effect on the longer-term dispositions, hence wUlbe most deeply involved with memory; they will play relatively persistentroles in determining how agents below them influence agents above them.In my image of development, the region in which agents are in this senseintermediate wiill presumably move upwards dnring cognitivp ~ r o w t h (See

Note 6.For example, a loW-level agent in the visual system wOl:lld always compute

. the same function of retinal stimUlation. But a t h i g h ~ r levels, differentdispositions would induce different ways of sE 0ing t h i n ~ s . 'For example,the choice of perspective for the Necker cube is dictated , not by ascendingsensory information but by preference signals coming from other agencies.Thus, one Uses non-sensory information to dispose oneself to rrp,ard sound asnoise or word -- or image as thing or picture. Each P-pyrarnid may have arepertory of such dispositions, defined by pre-activating different subsets ofagents. And a single such system might maintain, at one time, fragments ofseveral such dispositions -- but only i conflict;; are not too serious,

K LINES al d LEVEL BANDS

Now imagine the whole brain to include many SUGh P-s t ructuresinterconnected and overlapping according to Intricate genetic constraints.Return to the psychological view for a moment, and suppose that one part Pof your mind. has just experienced a mental event;EK which led toaChieving some goal -- call i t GK. Suppose another patt G of your minddeclares this to be memorable , We postulate that two things happen:

K-NODE ASSIGNMENT: A new agent -- call i t the K-node AK -- iscreated and somehow linked with GK.

K-LINE ATTACHMENT: Each K-node has a K line -- a wire havingpotential connections to every Agent in the P-pyramid. The act ofmemorizing t causes this K-line to make an excitatory attachment to

every currently active P-agent.The result: when AK is activated at a later timf), its K-line will make P

re-enact that partial state -- by arousing those p-agents that Were activeWhen EK was celebrated. So P will virtually hallucinate that event. (SeeNote 1.


9/24

Theory of Memorr

t l i l ~ i s l h e basic idea. Blltitmlght seem impractical because E, yery R-ltneh ~ { ) o c c : ) ' m e n e a r every P-agen t. We now in'trod uce a series ()f'impfovements that combine to form a powerful mechanism for abstraction

a tid ., .fitetence. First let us .note thClt i t is not the goal of Memory to, t 6 d u c ~ ' a perfect hallucillation. (See Note 8.) One wants to re-enaci only. enofighto get the Jdea", Indeed. the perfect hallucination would be.. hartf l tut tol completeiesetting of the P-net would erase all the work donein p r o ~ e S S i n g the recent data. It might even fool one into speing the' presentproblem as already solved. The new state must he sensitive to the newsit:uation.A memory should ,induce a state through which We see current. re'allty as ~ n instancl o th E? ~ e m e m b e r e d event. The idea below how todo Jhlsts probably the most important idea of this theory.

TilE E V E L ~ I J 1 l N D PRINCIPLE1'0' obtain the desired metaphorical activity. we do not connect AK to all the'p-ageilts t l iatwere active during EK. but onJy those within an intermediateba 140f levels. To explain nits, we must suppose that AK Is somehow

~ S o c l a t : e d with some agent PK at a certain level of the P-pyramid -- W edlscusstliis "P_:':>Ki , association later. Then:

.' t P . w ~ R : b A N t > - t . I M l t : the. K-line must not reach agents at levels' far:belowPK . for this would impos false perceptions and conceal ihe realdetai1.s of the present problem. (See Note 9,)

.. tJPPf:R BAND-LIMIT: Nor should that K-linn reach up close to the level',of PK, itself. for that would make us h a l i u c ~ n a t e t h . e presl'ntproblem 'asalreadY solved, and impose too strongly the d{ tails of the old solution.

. ' , . ' ' t /' '. . - . , .Thes.e ,two constraints combine to suggest:, 'LEVEL-BANI) PRINCiPLE: A K-line should span only a band 1 levels

~ d ' m e t t ; h e i e b e l i J w i h a f o { P K , leaving It ~ e e loCi) exploit higher ~ 1 t e lagJhts appropriate t6 current goals and '(it) be sensitive to currentcQntigenc1esas perceived at lower leveis. ' .

. . . , , ' .e,. ' . ' , , . . : . - -


10/24

1\I \

1 \I \

K P \too hi h I \I \ \ \

/ 1\ \1 > 1 > / \ \

level-band 1 1 1 I 1\ \1 > 1 > 1 I \ \

\ \ 1 > 1 \ \ \ \\ > 1 > 1 \1 \ \too low I 1\ 1\ 1\ \

/ I \1 \1 \ \I I 1\ \ 1\ \

I \To summarize: by activating agents only at intermediate levels" the systemcan perform a computation analogous to one from the memorable past. butsensitive to present goals and circumstances.

CONNECTIONS AMONG K NODESA second important principle is this: if K-lines are to contribute to memory.they may as well benefit from memory When forming a new K-node. weshouidhot ignore the existence of other, previolls1y ~ f i n e K-nodes. Heres how we embody this idea:

K-RECURSJON PRINCIPLE: Whenever you solve a problem, you exploitmembries from the past. So we can assume that when the memorable,event EK occurred, this itself was in large part due to activation ofsome already-existing K-lines. h e r e f o r ~ it will suffice to attach thnew K l ine AK to just the currently active K-nodes

In effect, this says that new memories are composed mainly of ingre lientsfrdm earlier memories. By making connections to other K-nodps (rather thanP:-nodes) we need fewer connections and obtain (we shall argue) morel tleaningful cognitive structures. The level-band arguments apply just asbefore, .hence:


11/24

Marvin Minsky 1 K-TheC n: of Memory

We do not connect AK to a l ~ the K-nodes active during F.f . but on ly tothose n accord wit the Level-Band PrincipleTaken literally. this has a fatal flaw: j K-lines contact on.1r othpr K-nodes,they can have no ultimate contact with the P-pyramid. That process has tostart somewhere Our proposal: we envision that K-agents } i f anatomicallynear the P-agents of corresponding levels. Then i t is easy for K-Unes tocontact either p- or K- agents. Presumably genetics specifies the proportionsand, during development, these preferences tend to shift over from P's to K's.

THE CROSS ILR PRO LEM

Evenushlg the Recursion and Level-Band principles, still pach K-node needspotential junctions with many agents. Every brain theory must deal in someway With this "crossbar" problem -- to make the mind capable of a greatrange of "associations". There Jnay be no general s o l u t ~ o n . In the cerebralcortex, for example, the (potential) interconnections constitute almost theelltire biomass and the actual computer is but a thin layer bordering a threedimensional mass of connecting fibres. Btl t not ' that the Level-bandprinciple does reduce by one the apparent dimenSionality of the problem.(See Note 10.)The crossbar issue is often ignored in traditional programming, becausecOI1l1niter memory can be regarded as totally-connected in thr:: sense thatregister "addresses" can connect any cell to any other in a single step. Theptobiem returns in systems with mtlltiple proces;,ors or more active kinds ofmemory.

In arty case, 1 would not seek to solve the crossbar problem within thec o n t ~ x t of K-theory nor, for that matter, in any clevPT coding scheme, a tholographic phase-detector -- although any such inventions might hp.lp makebrains more' efficient. Instead, I would seek the answer in the concept oftheS,()Ciety. of Mind itself. I f the mechanisms of thought can be dividedinto s p ~ c i a i i s t s that intercommunicate only sparsely, then the crossbarpfobl,em may need no general solution,: for most pairs of at,ents will haveno . teal needt talk to one another. Indeed, because they speak (so t ) speak)dlf 'terelltlanguages. they could not even understand each other. If most contmtlnication is local, the crossbar problem scales to more modest pf6t'orHons.The reader m i g h ~ complain that communication limits within the Mind seemcounter . n t u i t i v e ~different?

cannot one mentally associate any two ideas, however


12/24

Marvin Minsky 11

Though the final answer is surely "yes", i t would seem that unusualconnections are usually "indirect" -- be i t via words, imag( s, or whatever.The bizarre structures used by mnemonists (and, presumably unknowingly,y each of us) suggest that arbitrary connections require devious pathways.

TilE KNOWLEDGE TREEI t will not have escaped' the reader that we have arrived at an elegantgeometry:

I

ThaK-nodes grow in to a structure whose connections mirror those of theP-pyramld except that information f lows goes the other way Pno.dc:js activate units above them K-nodes activate unif[; below them.Thus forms a K-pyramid ly ing closely against the P-pyramid eachwit convenient access to the level bands of the other.

1\ \/ \ / \

/ \ / \K \ P \/ \\ \\ \/ KIP \ \

/ \ \I K P \ \

I I / \ \ ,if/ I \ \ \lI 1 \ \ t/ I / 1 \ \

I I I I > \ \I I \ \

/ I I>>>> >> \ \/ I I \ \

I I \ \. .in, t e r ,mso f th i sd i ag ram, the local pattern of computation, fornu. a

counterclockwise spiral. Globally, over several "cycles", the loctis of activity. can d d f l either. upwards or down. This "computational architecture seemsvery genetaland versatile.


13/24

~ 1 . i t , t h e apparent symmetry lsde,cejitive. beca ilse r sllPpressed some hardq t i e s t i ~ h s . t gave adequate descriptions 'ot t h e c o l 1 n u c t 1 o n s w H ~ i n K,'and ofthose' f ~ p m }{ to P. r sa,id little a ~ ~ l l t t.hE connecttons witnJ.n P, but that isnot p,art ot this story, nor is it a problem here; this Is disctlssedin [t].But of the cortn,E)ctions from. ? b , ~ c k to, K, all 1 si1fd was that ..... AK issomehow asSOCiated with some agent PK at a certain 'level of the, P-pyramid

f I '

The,tgeawas Some way to relate P-f,wtmts w H h a ~ h l e v > m e n t of Goals'represented elsewhere. ,The rest of the essay discusses various possible suchrelations but does not settle1,lpon any particular one. Itt fact, this ends theconslru'ctive part of this essay and, from this point, the reader can a s ~ t l t t t ethatditflc1,11tles in understanding are my fault, not his. I hope 0I11ythat

t h : ~ foregoing intuitions' may stimulate others, to constiuct a more completethebfY.' 'It:iS'tenipting to try to find simple ways to restore the symmetry. For

'e1Cample,we talked only o t m a k h l ~ f the K-treelearnto adapt to theP-tree.bllliite, P: .tree itself must once have been the l ~ a , r n e r . Could they taket9,tris,txaining each other? Was the P-tree once theK-tree for another p.:.

s y ~ t e D l . ?

'Mas,nothing so simple will do. We later argue that non-trivial learl1ingr e q u r r ~ s at least t11.ree nets to, be ' .inyolved.For there must be s o m ~ l inkfroni k. and P to the rest of the Society, and the P-->K connection seems towafifthili tole. K-KNOWLEDGE

, : , , , - . , , ,W ~ ~ t a r t e d w i ~ l 1 a n a i v ~ idea that memories re-enact past states wiUl.out

a t t e u t ~ t i ~ f g ' toexpla\n wh'a't they mean . But riow > we comP. ' full circl.e:$ t t t ( e t i 1 ~ K-systetn' (ormsa SPft of hieratchlcalwe\l,. one can hardly escapeas'ktfig what its nodes might mean. I t seems n.atural to try to see it as some , sor t h'r abstraction iattic::e tq which each K-node represents some relation. ain,ong W hatever its subprdlniites /I repiese.n t . ' . ,

, , . . . ,K-k4i>W 1,dge seen as r o g i c ~ 1 What' kinds of rf;'l


14/24

Marvin Minsky 13

, . I

When the partial states of the subordinates do interact, the "logic'; ofc o m b i n i n ~ K ..nnes depends upon the "logic" within p, In a vf)tsionofcrossexclusion that Papert and I favor, the activation of two or more competitiveP-units usually causes their entire closs-exclusion group simply to "dropout completely, defaulting to another gr()up at the next hill,ht'r level. Wesee this. s profound heuristicprincipJe: i f a single viewpoint producestwocori,flicting suggestions in a certain situation, i t is often better not toseek a compromise, but to seek anc;>ther, less ambiguous vif:'Wpoint -Weintroduced this, .idea as a general principle in [2] after Papert formulated i ts thaoryof how Piaget's Conservation develops in Children,

J{-K,nowledge seen as Abstract. Initially, W(, spoke only of creating anafitirely new K,-node for each memorable event. Now we bf'gin to allow formbregradual and incrE.'mental ways to "accumulilte" new suborrtinates to anexisting node. A ch.impanzee might reach the too-hir,h banana by different:means atdtfferent times -- first using a box, t h ~ n t chair, later a table. faU th,E $E can be "accumulated" to the same node, it can becomp. a poWerful'ihow, to reach higher" node. When re-activated, i t will concurrentlyactfvite I-agents for boxes, chairs, or tables, so that perception of . . l ~ ofthem will be con. 'iidered relevant to the "reach higher" goal. In tilis crudeWi lY,s'Uch an "accumulating" K-node will acquire the effect of a class

~ b . s t t ~ e t l o n -- an extensional definition of "something to stand Oil", ':Btltft . inlY dpmuch better than that If conflicts betweendetaHs' cancel o m ~a tofhetout (because o f conflict within cross-exclusion subgroups, as

'. mentioned' above) then d ~ c i s i o n s wil l default to the remaining nonconfiicting details ' This automatically prodilces a more abf;tract k ind 01abstraction -- the extraction of common, IJon-conflictingpropertiesCciIl1 t:,fnhtg the concrete "accumUlation" of particular instances with the

r ~ J e c H o n o f strongly dissonant properties leads automatically to a ratherabstract 'unification". (See Note 11.) .. ',,'

'k-,1tll(,",, edge as Pl'ocedural. When K-lines interact at different verticall ~ v e l s ~ t h e superpOsition of several partial states will produce various sorts of. . l b g i c ~ i .and"illogical consequences" of them. We already know they can.pr',odbct:\ shnple disjuncts as well ,as "exClusive-ors" enollr,h to make atl b;hietsa1 propositional logic. For ,predicate logic, a lower ,K-line could affect. he i h s t ~ n . d a t i o r i of a higher-level, "more abstract" K-Hne,' For example. this'bdbld1:lea way to partly instantiate one frame [3] with other frames at itst ~ t m i n a l s Thus, group of K-lines could activate a frame displacing someor its I'defattlts assignments" by active sensory recor,nizers.(See Note 9,


15/24

Marvin Minsky 14

What else can happen depends on the specifics of the P-lor,ic. One mightevert be able to design a "detachment" operation to yield deduction chainingvia the overall K-P-K-- operation cycle. Bu t I have no detailpd proposalabout how to do that.

LEARNING and REINFORCEMENTGenetations of experiments have led to many theories about learning -- inaniIrtals -- via reinforcement of SUCcess But I.maintain that no suchshrtplistic, centralized reward mechanism could suffice for human learningbecause: .

The r e c o g n i ~ i o n of what eventslShould be co nsidered ./memorable -- nail inte l l igent system -- cannot be a single. uniform process. trequires too much intelligence. For the purposes of any particulardIvision of;the mind, such recognitions must usually be made y someother agencY that has engaged the present one for a purpose.

To solve hard problems, one needs strategies and tactics that span verydiffetent ' time scales. When a goal is finally achieved, one wants to

reinforce not only. the immediately preceding events, but al:;o the longerrange strategy that caused them. But, between sp.lpction and completion of astrategiC plan, there usually intervene a variety of tactical failures. So atthat final moment the traces that remain within the mind's state include allsorts of elements left over from bad decisions and futileexpertments.Traditional behavioristic learning theories rely on "recency" to sort these out.This could work for simple strategies in which the most recent mentalevents are indeed the best correlated with sucCPss. But for human problemsolving. I am sure that the credit assignm('nt problem is much toocomplicated for this to work. Instead, I conjecture, different scales ofstrategies and tactics are segregated in different agencies -- e.g., different Pnets. Then, the learning mechanisms can also be segregated to operate overdifferent time scales.. After. all ,many human cognitive strategies actuallyachieyE their goals by assembling subsidiary learning systems that operatebverhUtirs and days. Strategies for dealing with Joss and grief, acqUisitionattd ambition, span the years -- yet, som.etimes. in the end we learnsometh:t.ng from them


16/24

F o r ~ h a t . that is worth, we conclude that decisions abo\lt what and when toreinforcefi cannot be made on a global recency basis., Nor can it be done

eniitely locally within the K ~ P pair, for they lack eno1.tgh information about the intentions of other centers. At the least, i t would seem that c(mttot oyer formation of K-P links must be held by a third agQncy -- either onewith iItnate, unlearned reinforcers, or one that has alrf.'ady learnedsori let l1ing .C0rtsiderarnodel based on these intuitions. in which a thhdrietwork a,WJth.2lIl ~ c t i v e goal-node GK, has the power to construct new K:"nodes for P.:ttPPdse that at some earlier time UK was achieved and was connected to at{';;nodeKG that activates tWo subnodes K 1 and. K2. At some later time Gachieyesanother instance of GK and celebrates this as r n e m o r a ~ l e . I f nothingnErWhappened in P, there is no need. to change KG. But Slipposea new

e l e r f i ~ n t K3...;- P3 1s involved this time: then we could add :K to KG's Kl i i te,so th.at P3w1ll be available for achieving GK in the fu:huE'.

\GI \ < - - - - - - - - - - - . - ~ - II \ I

I K \ II \\ I

I \ \ I

II)

.\ II \ III \ I \P

K K\:;\ \/ 1 \ I \

I Kl--\---Pl \K 2 I ~ P 2 \

IK3-1 ---- --P3 \ \OfcdUtse, t.his raises all t.he issues about novelty,: conflict, adaptation ands a t t t f ~ H o : r l t h a t any learnIng theory must face. (See Note 12. What 1fP3l a l e rbecan i eadhec t competitor of l o t 1 21. What if there Wert 8mistake? How do we keep the web attached to KG witllin bounds? Afterall. there is always som thing new One can try to invent local solutionstoal l th t l se problems, but I doubt there is anY' s i n g l e . ~ . a d e q t i a t e answer.Instead, i t must be better always to feave fink. formation u r l l f ( ~ r the cOntrol

o t a distinct system that ifse1f can learn, so that themne:monic strategIes .in e ~ h l ) c a l e can be made to suit their citcnmstances. ~ e ) : h a p s some people'become. stnarter than others . because they d:evelop better nnH lttonlc strategies. 1 h ~ ~ ~ . I 1 i t g h t more affect the quality of intelligence t h ~ n do the specfficpl'ohletn';;solving strategies we can observe directly.


17/24

Ret1.ltn)ng to the three-part m o d e l w h ~ t activates KG? I f the G-systemc ~ t 1 i d . c a l l on a variety of P-nets for its p u r p o ~ e s GK might be selectedbecause of some cue involving P that suggests i t as a plausible alternat.ive.....e.g KG is activated by an and of OK a n ~ d that P-condition. ThroughSl1c:h COIUtectiOl1S KG becomes part of the representation or meani'ng of GK -a remembeted solutiol;lto a problem. While this roisE s more questions thant answers, i t seems clear that a minimal learning thl'ory will involve at

least thr nets -.,. G, K, and P -- in which the first controls how thesecond learns to operate the third. This does not mean the system is madeQrdistinct such triplets. Presumably, the same net could play a P-role inone d.omain and a G-role in another.

T CIT VB RTICUL TE KNOWLEDGE ..

t is commonplace to distinguish between "tacit" knowledge (like how toclimb stairs) and expliCit knowledge (like how to spell "spell"). In asil1g1e-8gEmt theory, one might wonder how knowledge could possibly betacH. In a "society of mind" theory. one mit,ht wonder how couldkit6\o'{ledge ever become explicit . One cannot expect positive answers ingeneral; only where K-->P connections become somehow linked with suchc6gnftbre elements as particular senses oJ particular words.t t i s be t t e r to regard the "tacit-explicit" distinction as. merely a firstappr6xlmation to some richer theory of thediffe,rent: kinds of remoteness

~ e t ~ ~ e n o n e mechanisnland another. While surely someagendes in themind. have exceptional expressive roles,each sub-society of the mind must s l l l lhave its own internal epistemology and phenomenology, with most

details private, not only from those central processes, but from one another.hfmyvtew self-awareness is a complex, constructed illusion. As. adults wetightly place high value' on the work of those mental agencips that acquirep6\.Vers to reflect on the behavior of ot ter agencies -- especially out

l i n g i i ~ i s t l and ego-structure mechanisms. But probably no part of any mindc a h ~ v e t see very deeply into other parts; t can only use tllodels i t. construCts of the.m. Any theory of intelligence must eventually explain the. agencies that make models of others: s ~ h self -awareness is probablyessential to highly intelligent thought, because thinkers must adapt theirstrategies to the available mental resources.Each part of the mind seeS only the surface products of some other parts.What little we can "directly" sense is swiftly refined, reformulated and"represented", We find i t useful to believe that these fragments have'


18/24

..,

, a r v i n M l t 1 s k Y 17 . K - t h ~ " o J _ of Memory

; i ~ ~ : r i . J . n g s in themselves apart from thlf grMt Websbt structure from whicht h ~ Y e m e t g e . That illusion (valuabletopi: ople qua thinkers but not quapsycholqgtsts) leads us to think that e pressiblc knowledge iB the (irst thingto study. H the present theory Is right, this is topsy-turvy; mostknowiedgestays more or less Where nwas formed. and does its work there.i t l s the exception, not the rule, that lets OM s p ~ a k of what one knoWs

. Tosaymuth mote about this would engage a world of issues beyond thebottnds of this little t h e ~ r y . 1 mean to indicate no pessimism in saying thatexplaining the meanings of membries will need many more Httle theoriesbeYO:hd thi.s OI\e. We can understand the ,jmeanings in t h ~ partsoC our.l11.inds .1t .. andonly i f ... we can model en()ugh 'of them inside others Butthis is no different from understanding anything else ' :except perhapsharder. . . . .

NOT S

. Cambridge, MassachusettS"January - June, .1919

, :-l ~ r ~ t ~ r u l t : y a C 1 t r i 6 W l e d g e \ ~ l u a ' b l e ~ f s c t i s , s l o r i $ aboutK-Unes with' n. Hillis, G.

J ; : ~ h i s s m a n , W. Richards,' Jon Doyle,R. J. Solomon.off,lt 'BerWick, and,espedally S. Papert - for the bastc Idea carne in conversations with him.

oter : Backgl bUhd . The references to the "Societyof MiItd"relate to a. t l i e d ~ Y l h a v ( ; been eVolVing jointly wifhS. Papert. That theory 'tries to

" ' e ) t : P i ~ i n J h ( ) \ 1 g h t in t ~ r m s o f many weakly interading(and ottctt ctinflicting)speciailsts,father than in terms of a ce'ntraHzed and logically, consistent, s y s t ~ t b . ' I t is described briefly in [1]. which the present papetcomplements

' i n . s ~ v ~ r a l .arMS. The. c'omputational s ~ f u c t t t t f ' S described ,therein Were. ( ; o h : f t t ~ U ' l g l y b i d l r : e e t i o l 1 a l l and ,the K-P dttaiity c1arifies that a little. The C-, i1lttls ,ofttiatpaper correspond roughly to the g >p connections here. . The

~ . ~ $ ~ t t $ ' s l ( ) n i ~ [ tJ of. cognitiv e ca:u:ls and ofdifferencessupplemEmtthed l ; s ~ l i s s i o n hi;lie: of goals. But 1 do n.ot mean to pretend that the reader. $hduld be able to figure out, even from both papt'ts, exSctlywhat happensin P-fiets; We simply haven't fixed the details.


19/24

;:.:.M:=;;a:.:::r...:.v1-=.;n==- ..::M=ln=.:s::.:::k::lly - . : l : . . . 8 ~ - K : . . : : - . . . . ; T : . . . : h = e o . ~ y of Memo >:

Note 2: Disposit ions The term "disposition" is used here in its ordinarylanguage sense to mean a momentary range of possible behaviors". I don'tsee a way to define t technically without making t synonymous withstate , which do.es not capture the same intui tiofl . In a computer program,a disposition could be imposed by selecting which items are active in a data

base, e.g., as in Doyle's [4] flagging of items that are in and out ofcurrent consideration.The term representation also has problems. t always involves three agents-- A represents as C. In a Mind theory, A might be either part of themind or the theorist himself; one must be clear about. that In this paper, aK-node iIilposes a disposition on a P-net hence. o r us that node can

represent that disposition. But what t represents for the mind that containsi t is another matter we touch on only at the VNy end of the pilper.Note 3: Modularity This i5not to say that understanding memories offeelings should be easier than understanding memories of facts. The lattera.ppe'at simpler in the adult perspective of "modular" knowlc(lge, because al i fet ime. of mental theory-construction builds for us our order ly ,cotnttionsense epistemological hierarchies. A fragment of incrementalk.ri'owledge -- e.g., that ducks have webbed feet -- is easy to repre5ent ,once we have only to link together a few already established structures.But should not mistake that surface smoothness for simplicity of underlyingprocess. It captures little of the real quality of "ml:ianing" -- of how suchlinkages participate in the total "web" of our dispositions.Note 4: Brains Some might object that we jllst don't know enough aboutbrains to make such theories But we are not propo5ing specific neurological. details only that things are or(,anized along the general Hnes of theSociety theory. This architectural theory is just another form of informationprocessing theory, emphasizing control structure and data flow rather thandata structure.Note 5: UnidirectionaIity t is technically very difficult to make theoriesabout systems thar allow large degrees of circular hehavior. On the otherhand. one cannot base a theory of mind on unidirectional netyvorks, becauseloops artd feedback are essential for non-trivialllehavior. This is why t has~ n sO difficult to pursue the field neural net models, and why solittle has happened therein since the works of Hebb [5] and M.ut [6].


20/24

MarVin. Minsky 19 K-Theory of Memory:

What J f ind satisfying is the way the p r ~ s e n t theory introduces the rf.:'qulredcircularity in a c o n ~ r o l l e d way. t hegins with a nearly uniditectionalnetwork,avoiding 0 mathematical universality and its usual theoreticalintractability. (The lateral cross-exclusion of the P-nets st i l lkaves basicallyunidirectional behavior.) Then, feed hack loops are built II p as steps intrainin.g the K-net. Surely this strategy lends itself to circuits that aremanageable and 'delmggable. With the loops intrOduced a little at a time, onecan watch for instability and oscillation, distraction and obsession.'We note that K-logic must be more complex than as described above. factivating a K-node recursively activates subordinates all the way doWn, this

w ) u ~ d v i t i a t ~ the level-band idea. J do not see any e ~ s y local way to dealwtththi l5; i t suggests that the activity band of a K-P pair should becontl:'.olled, not locally, but by some other agency - using a facilitationsignal (with low spatial resolution) that enhances the activity in a selectedlevel band. Such an agency could bias the a >cent or d e s c ~ > n t of the K-Pcomputation, without needing to understand much of the details of theeve.nts within K-P. In effect i t could instruct K-P to try a more generalmethod or to pay more attention to the input or, perhaps, to try anotherUke thi ft , and so forth.Such an agency would provide a locus for high-level heuristic knowledgea 'bbuthow to use the knowledge within K-P, and would be useful foritnplementingplans, looking ahead, and backing up It might he the naturalplace for OUr .all-important knowledge about knowledge .More speculatively. perhaps the difficulty of dealing with too-circular

ne tworkS is no mere human limitation. Evolution itself probably cannotcope.with the uncontrolled range of recursive network behaViors. So, we. speculate, the individual nervous system h s to evolve its circularities byseparating the flows into distinct d i r ~ c t i o n a ) cIa.sses. f the present theoryWere cOrrect, this would suggest an evolutionary pressure that might haveled to it.Note 6: Global Architecture n entire brain would contain many such Pstructures associated with different, genetically specified functions: sensory.motor,affective, motivational, and whatever. the present theory wouldapply only to the common properties of neocortex: the brain contains manyother klnds of structures. Finally, I repeat that the pyramid image is onlyrelative toariy particular agent . there is no reason ,to suppose that Pst.J:'Uchires need either narrow or widen as they ascend.


21/24

Marv.in IVUflsky 20 K-Tl:tf.'qty of lVlemory

; .

N:ote 7: Exc;itation. G.A. Miller pointed out to me that this resembles the. idea of retiintegratton . popular in an earlier era of p s y c h o l o ~ y . Note thatwe do Bot need add . negative;' K-line' connee'Honsto agents that Wereinactive when EK occ:urred; many of them will be automatically suppressed. by cross-;exClusion via AK. Others may persist, so that the partial'h;,lllttcbiat:lon may include additional e l ~ . m e p t s . I t is perhaps of interest that(according to Mountcastle [7]) all lines entering' the cortex from other.canters are excitatory.

N o t e ' ~ : A c ~ l l , . a c y . Only a naive theory of memory would df.'pend crfticalty. onrfrs};.Ume perfect recollection. Many agents active durhig EK wiU be ' t n ~ s s j ~ r i . H a J ' ; t o most new Situations, so we need not demand perfect andcompIet.e a t t ~ c h m e n t s ; indeed we will need ways to Correct serious errors. ' t ~ t e r . ill ihe eariy days oC'simple neural models 'one might h ~ v e welcomedsa'mpiing noise as a d e ~ . r a b l e . source of variety. In systems as' cotnplicatedasth.e p t ~ s e n t one; that' view is obsolete; the problem is, rather,of flndiilgheuHstlcs to restrict excessive variation.

,'. .1 V 9 t e 9 ~ F : t i l J g ~ s aHdFrames. n this sense, a K-node a c t ~ l tkea f r ~ m e ; as.. d ~ s c t i b e d in. [3]. When a K-node activates agf'nts in the I p v ~ J - b a n d below

. .- it','t,hese. porrespbnd to the essential, obligatory terminals of the frame. Bym ~ k l n g k- l ineshave weaker connections at ts loWer frlnj:(?s, we obtain, ::; ...._,.: l- .,;;,.. .. . . : . .,.'. , ': , ., . , . 'c '.. , muchof.the effect of the loosely bound default assignments of the f rame

: . ; ~ ~ ~ : : , : ~ . ~ . . :' ; . i , .. ,.,. t , .- ', , ' : ...... .. . t h ~ o t y ~ For, ' ~ a k l y .activated agents will. bci less persi tE.?l1t in. ctoss-... , ~ , , ~ i ) i ~ i p I 1 : , c ' p n i p e t 1 t * ~ n . What about the .uppe'r t r i ~ g e ? T h i s mieht be reiatedt t ) t h ~ 'coniPlementary concept of aUframe-system , emphasized in [3J .AtaUttre'Qf t)l.ep,:,net to do anything useful could c a u ~ e i t to d ~ f a U : l t controlto a:.sltglliiyhighar: leVel goal type; that is, to move up in abstraction level.. Mltitis , c ) I t 1 e ~ simply from making weak connections at the fringes of the, l e v ~ f : b a n d .. ;, .-

t i e e ~ ~ r i . l z e t : t l a t t h i s argument about the u p p ~ r limit is much less clear thEmf d r . t h ~ lower 1 1 m i t ~ I have 110, realiy strong reason ' even to' Insist that', the, , p : j J ~ r r r l n g e end below N{, except for an overall feeling of cQnsistem;y. In

. r ~ f ' ~ ~ P K ~ ' W : a s r i ' t d e f i n e d dearly in the t i r ~ t p l a l : e , except for l ~ c ; a t ' i n g the........' l ~ ~ ~ ~ l l . , b ~ , ~ ~ s ..... But the, a s y m r n e t ~ r r e a l ~ y c o ~ e s ' from t ~ e m u r k . n e ~ ~ o ( , m y.' . . . . e x p t ~ l l a t I o n ( J r l towthe. P - - > K c o n . n e c t l o n ~ . r e l a t e P-structures. to goals and. ~ # J i ~ ~ ~ < ~ A,t the end of the. paper are a e ~ j n c Q ~ p l e t ~ suggestion.s aboutthe$em:aUers. .

,.' ,,;. ;, ; : -

' , , ' .

\ . . ' ; ~ ' : ;: :.'


22/24

M a r ~ i n M t n ~ k Y 21 K-Theory of Mem ?. .l

ote 10: Crossbar Problem I, am not very much concerned about thisp : r o b l ~ J l 1 t s s i z ~ , because I enVision the minI as employing a few thousand. p

n e , t s ; ~ a c h with a few thousand. Agents. So the local crossbar problem,whIch is the one' that 'concerns ~ most, 'involvfls orily t h o u ~ a n d s of lines,notmilHons; , that is, K-lines must have access to that order of connections.

s fo t interconnecting all the P-nets, this must 'be the function of thebrain's white matter; we argue in [1] that one ]lped not suppose all P-netscaRot need to communicate with each other.

,there eXist communication-hardware schemes more physically' efficient thanpoint; i6 point wiring. Since' the density of actual K":Une connections is, stitely s p ~ r s e in the space of aU possible such connecUonll, they could usesuch schemes as those of Mooers [81 or Willshaw et t [9). To implementthese, o n e w o ~ l d 1 t i a k e a v a i l a b l e a large bundle of descending conductors -caUthemM-lines. To simulate a K-Une, attach' the K-node to excite astJtall, .fixed, but randomly assigned, subset of M-lines. to connect t to

'an.other K;'",Qde, the latter must first construct the corresponding , logical,Midi., then "iogicai or" that into its excitation condition. lYsing to-linesubsets of a .100";Unebundle would suffice for very large K-pyramids. TheflnaLchapter of Fahlm.afi',s thesis [10] speculates on other radical crossbarschemes.

o t ~ l l :tvJnston Learning Because I consiler Winston's [11] the most, , biiert#s,tit(g constnictive theory of abstraction, 1 will tty to re1ate t to the',,;present theory.'itmphasis links" can be identified with K-lines to membersofd:oss,;,:exclusion groups. But "prevention pointers'" must enable specific 'PagentS.to disable higher level class-accepting agent; 1 do not see any easy

':W3f ,to do that. CruCial to Winston's scheme is the d(tection, and anaiysls ofpifferejJ.ces,To make our system able to do this, Ohe might wartt K-Une'a,ttatih.rti.entto prefer p-agents whose activation status has recently changed;th


23/24

~

'

Wniston s c h e m e ~ h i p h a s i z e s differences hi "near miss" sit iati oris; in a realSitUation there must be away to protect the agents from dissolution byresponding too actively to "far misses". Perhaps' a broader form of crossexduslon could sepa:rate the different senses of a concept into families.When serious conflicts result from. a far miss", this should disable the, cbnfrl:sedP-net so that a different version of the concept can' be formed .in

a t r o t h ~ ' P-net.N ~ t ~ , ~ t z : S ~ t l 1 r a . t i o l l , h l t f t e present theory, , o ~ e only addscontleetions. and

n e v e - r r ~ l t t o v ~ s tliem. . r h i s might lead tb'trouble. tloesa' perso l have a wayto , i ~ e d H l t b r prune his 'Cqgnitive networks? Well, the present theory. likeanytithetsimple psych,ologtcal theory, must proceed through stages -- just

'as:does its subject matter. Perhaps theWitiston theory could be aJllendedsot l ~ ~ t o n l ' y imperative :polnters long survive. Pfirh,a'ps the croSs exClusionJ b . e ( ; h C l ~ ( s m ' is adequate torerar lOW-level confusions to htghfJr level: agents.P ~ i r l 1 a p , ( when an area becomes muddled and u ~ r e l i a l J l e . we replace i t by.. a:n:other':'-perhaps using' a speCial revision 'mechanlsni,. Perhaps in this sense

~ ~ a r e lilt. ilke Jhe Immortal' people in Arthur Clarke's novel [ t 2]. who from. tittle, to thne erase their . least welcome rt colledions.

,, ,; .

:

. ...... .< \

.,;

: : ,

.

:;- j :


24/24

.. , ~ . , '

Marvin Minsky.......____________ -:2:::,;3:::--,.--.,.-_ _ ~ . : : . : K : . . _ - T . : . . : : h : . : , . : { ~ r y of e m o r ~

REfERENCES. . . , .... . ... , :. .. ' h i M i ~ ~ k . Y , M . ~ PlalnTalk,about Neurodevelopmental Epistemology'*, IJCAI,'1971. COJU1EmSed "e,rsion in Win,stonand Brown (eds.>, Art i f ic ia l

In.telligence, vol. 1, MIT Press, 1979[ i ] ~ and S. Papert,Artificiallntelligence, Univ. ' of Oregon Pless, 1974. '[3] MiIlsky, M., "A trCimework for ~ e p r e s e n U n g Knowledge'''tn Winston, ( ~ d . ) Psychologyo( Computer Vision, McOraw-iHll,1975 '

, . [41noYle, J., A Truth Maintenance System M1T AI Memo 52t, 1919:[sl.Hl: bb f D.O., Organization of Behavior, Wiley, 1949. t ~ f ~ ~ t ' r , hA Theory of 'Cerebellar cqrtei , Physiology,vol, 202,pp431;.47Q , 1969. Also, IIATheory for C e ' r e b r ~ l Neocortf. X", Proc.:

, ~ j y , S o c . Lond., vol. ~ 7 a B , ppt61-234, 1970'ftl Mountcastle, V., in The Mindful ~ a i n F . SChmttt (ed.), MtT Press,1978

r ~ r ' 1 \ ~ e r s , c . t . , Z a t o c o d i ~ ap.d DeveloJlments inInformation Retrieval",;>. A.SLIB proceedings, vol. 8, no. 1, pp3-Z2, Feb, 1956. ' , : y , . . , , I, [91: W.hlshaVi.P.J B u n e I 1 l a n ~ O . P ~ and H.C;J.,ong.llef-Higgtns, :'Non-::, Hblbgraphfe Associative Memory,Nature,' \ToL ,222, pp960-962, 1969.': .

t i b l X ~ b i ~ a n , 5., NETti a Systemfor RepreselltJng and {/sJngReaJ-WorM, ~ : i i ( ) w i e d g e , MIT Press, .1919 .< : ; ' . ;. , .:, '.' : ' ..... . . . . . . : . ' , . ., '. . . : ,"' .'. . , ''[11] Winston,P., "Learning StructurCll De:;;cripUons f r o m ~ x a m p l e s j in" . Wlnstpn(E d.), Psyc;hologyof CompriterV ision, McGraw Hill, 1975

, [l2JCtarke, A . C . ~ the City-and the Stars; Sig'net,1957

Documents

A Theory of Memory-Minsky