c
----- v,
-----v,
-----v,
Figure 5: How the TSP works. Nine cells wit h three V regist ers
are shown. Bar graphs represent t he symbol concent rat ion after the
sequence addefeais learned. An act ivated cell traps the current symbol
concent rat ion in one of it s V regist ers. For example, the first d traps
only some G
a
in its VI regist er. Each regist er represents one vector
that can act ivate (fire) the cell.
is dist ribut ed t hroughout t he volume, where it can be associated to the suc-
cessor symbols. Lat er , dur ing t esting to det ect t he existence of a stored se-
quence, t he sequence of emission and react ion-diffusion" t hat occurred during
training will be recapit ulated, exact ly or app roximately (depe nding on as),
by guided sequence retrieval.
4.2 An example of the t raining algorithm in use
An example of the st ored V vectors t hat would exist after a sequence is stored
is shown in Fi gure 5. The sequence addefea.. ., is stored.
When a first occurs, the environment is bl ank, so a act ivates Cell "a"
(through the ext ernal Z inputs, not shown), which record s as, the cur rent
local symbol concent rat ion, in it s first V regist er.
When the first d occurs , the symbol a has diffused to the vicinity of cell
"d," whi ch becomes act ivate d from t he external inputs through it s Z weight
vector. Being t he AC, it sto res the local symbol vector, G , in it s VI regist er ,
consisting onl y of some a symbol concentration.
When the second d occur s, cell "d" is reactivated and again st ores t he
local G , whi ch contains some a and some d symbol concent rat ions . Cell "d"
then emits a uni t impulse of d symbol.
Upon reaching the second occur rence of a in the input sequence, the
cur rent symbol concent ration at the x , y positi on of cell "a" is stored in the
second V regist er of the cell, which t hen retains t he state. The second V
regist er of cell "a" shows the successively smaller values for e, I, d, and
a, which are approximately propor ti onal (in this case ) to t he recency of
occurrence of the symbols in t he input sequence.
4Not e t hat symbol concent rations may grow as well as decay, depending on values of
the free par ameters in equations (1) and (2). This allows mor e degrees of freedom t han
diffusion alone.
316 Sylvian R. Ray and Hillol Kargupta
The actual magnitudes of t he symbols in the symbol vect or , G (x , y),
depend on t he relat ive posit ion of cells in the volume as well as on the
const ants used in t he reaction-diffusion equations . The magnitudes of symbol
concent rat ions do not even have to decay monotonically; it is essent ial only
t hat the same subsequence produce t he same concentrations on repeat ed
appearance .
St orage wit h as = 0 is analogous t o eideti c memory and requires a dif-
ferent V regist er for every uni que symbol transit ion. This condit ion is very
profligate wit h memory usage. For increasi ng as, however , an increasing
number of t ransit ions are t reated as equivalent by using an exist ing V vect or
t hat is close enough t o a subsequence already seen. In t his case , a new V
vect or is not required , and t his reduces storage space.
We have chosen t o use a one-shot t raining algorit hm here. Increment al
t raining, typically used in neural net works, could be used to average out
variat ions in timing of the symbols, but we will not pursue t hat avenue in
the current invest igat ion .
4.3 Demonstration of depth and counting quality
The depth and count ing ability of t he TSP, referr ed to in sect ion 2.1, was
demonstr ated in a simple experiment simulating only the st orage algorit hm.
This proper ty was demonstr ated by monit oring t he growt h of V regist er usage
while storing a sequence wit h repet it ions of a single symbol. The experiment
consisted of storing t he sequence ddddddd. . . and noting the symbol number ,
n, where the storage algorit hm first did not invoke a new V regist er to store
t he state . This would correspond to a depth of n - 1 as well as the ability
t o pre dict the different success or symbol for t he sequences cdn- 1e and cd" t ,
for example.
This experiment was performe d wit h various values of tolerance, as. The
result s are t ab ulated below. Unless ot herwise not ed, experiment s were run
with D
g
= 0.3, Dr = 0.3, a = (3 = 0.3, E = 0.2, and I = 0.1 in equations 1
and 2.
00
Counting Test
n (lar gest depth)
3
4
6
10
0.05
0.04
0.02
0.01
o
When t he recognit ion (or ret rieval) algorit hm, discussed lat er , uses the
same value of a , t he storage pr ocess is recapitulat ed determinist ically, so we
can conclude t hat t he depth during retrieval will be t he same as t hat found
by t his experiment .
A TSP Based on the Biological Reacti on-diffusion Process 317
Met aphorically, lar ger values of o, corres pond to paying less attent ion t o
the pr ecise count . As the needed count or depth increases, t he penal ty is
an increase in the required st orage capacity, the t ot al number of occupied V
regist ers in t he TSP.
4.4 The network in the embedded sequence recognition mode
Suppose many short sequences are stored in a ReDi TSP (as in Figure 1) . Let
a long argument sequence be pr esent ed beginning at it s kt h symbol , Sk , for
t he purpose of det ecting whet her t he arg ument subsequence is st ored in the
network. This is guided sequence retrieval ; we are using t he known exte rnal
sequence to guide t he search for it s stored equivalent .
It is assumed that (1) we want to search for the ent ire sto red (short) st ring
and (2) the period of symbol pr esentati on is t he same as dur ing t he storage
operation. (JT is t he tolerance measure during tes t operations , set t ing t he
radius of acceptance of mat ches bet ween the stored and argument sequences .
The algorit hm is as follows.
Guided Sequence R etrieval
1. Set the argument st ring point er , k = 1. Clear all symbol
concent rat ions , G( x , y), to zero.
2. Using the exte rnal input Sk, act ivate cell C
S k
(whose lab el is
GSk ) . If none, t erminate wit h failure.
If t here exists a V regist er of C
S k
such that
I: I (Gi(x,y) - Vi) IS; (JT
then C
S k
becomes t he new AC and the t ransition Sk-1 --> Sk
in the argument sequence is a known (previously encoun-
tered) transition. Emit one uni t of Sk at t he locati on of C
S k
,
and perform one symbol step of the reacti on-diffusion equa-
ti ons. Else, if no cell sat isfies inequalit y 2, t hen t he external
sequence is known onl y from S l to Sk- 1 . Terminate t he t est
wit h unknown sequence.
3. If a known t ransit ion occurred, increment k and return to
st ep 2, it erat ing until an end mark of th e stored sequence is
encount ered, imp lying a known sequence.
The foregoing algorit hm of t he t est phase recapitulat es t he training steps
in the sense that t he symbol condit ions present at st orage are revivified dur -
ing the successful matching of the argument sequence to an originally stored
sequence . Not ice, however , t hat we are permitting a different act ivat ion tol-
erance measure, (JT , than we used during training. As (JT ri ses from 0, we
are requiring increasingly less st rict mat ching of the symbol concent rat ions
wit h the ori ginal condit ions during st orage of t he sequence. The int eresting
feature here is t hat t he pr ecision of mat ching the external sequence t o the
internal sequence is controllable during recogni tion.
318
TRANSITION DIAGR AM FOR
REBER GRAMMAR
Sylvian R. Ray and Hillal Kargupta
EMBEDDED REBER
GRAMMAR
Figur e 6: Embedded rebel' grammar. Only t he penultimat e symbol is
predict able, given the second symbol.
4.5 An embed d ed sequence recognition experiment
Embedded sequence recogni tion provides a t est bed to demonstr at e how equiv-
alence class flexibility and storage efficiency are t raded off.
A finit e-st at e Reber grammar has been st rengt hened as a test for depth
and time-series predicti on in recurr ent neur al nets . The generated sequences
all have t he second and penult imate symbols in det ermini st ic relat ionship,
generating sequences such as BTXPSSTTTXTE or BPVTPSPE (Figure 6)
[8] . All interior sequence positi ons, other t han the penult imate, correspond
to transit ions having equal probability of eit her of two symbols.
The average sentence length generated is about 10 symbols, with t he
maximum of about 30 symbols in a sample of 500 sent ences.
A ReDi TSP wit h 25 cells in a 5 x 5 grid was defined. Five hundred
st rings were generated and applied to t he TSP in t he training or storage
mode. Since t he grammar uses only 7 symbols, t here were 18 unused cells.
Varying t he st orage tolerance measure, (J., demonst rat es t he range of
generalization possible during storage. In t he limit as a, -> 0, every t ransi-
t ion with a unique hist ory is stored as a unique transiti on, consuming one
V regist er. The first appearance of a distinct t ransit ion is signaled as an
unknown t ransit ion, which requires recrui t ment of a V regist er. As (J s in-
creases, some transitions are t reated as equivalent by t he system. For exam-
ple, when (J = 0.05, bpptttvvpe is stored as equivalent to bppttvvpe. Higher
values of (J s result in generalizat ion or enlargement of equivalence class, even-
t ually to t he point th at every t ransit ion is cont ext-free.
Similarl y, increased values of act ivat ion tolerance dur ing retrieval, (JT ,
permit more t olerance in t he affirmat ive decision t hat a part icular t rans it ion
wit h it s (possibly) extensive history is permissible.
The accuracy as a functi on of (JT is plotted in Fi gur e 7 for (J s fixed at
0.0001. Cases where a, > (J T have relatively lit tl e meaning.
A TSP Based on the Biological Reaction-diffusion Process
99 ~ ~ ~ ~ ~ ~ ~ ~ ~ ~
319
98
97
96
95
94
93
92
91 L-_--1::__L-_---l..__.L-_---l..__.L-_----'-_ _ -L-_--l__
o 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.009 0.01
sigma T
Fi gur e 7: Accuracy of pr edi cting embe dded Reb er gr ammar sentences
as a fun cti on of UT , the t est phase tolerance (us = 0.0001) .
600
0
Number 500
of
V-Regis.
400
Used
300 0
0
200
0
100
0
.005 .01 .05
Sigma
(storage activation tolerance)
Fi gure 8: Number of activation condition regist ers (V regist er s) used
as a function of us' Symbol tran sit ions tot aled 4492.
To ret rieve a sequence , a randomly gener ate d Reber sente nce is pr esented
symbol-sequent ially, and mat ching is at t empted by the guided sequence re-
trieval algorithm. The retrieval is counted as accura te only if t he penultimate
symbol transition in the sentence is found t o be a known t ra nsit ion." The
number of act ivation condit ion regist ers (V regist ers) required as a functi on
of Us is displ ayed in Figure 8. This number is a functi on only of st orage
t olerance, u" not retrieval t oleran ce.
5The a priori probability is 0.50.
320 Syivian R. Ray and Hillai Kargupta
Table 1: False positive rat e (percent) for recognit ion of embedded
Reber grammar sentences.
(58 I (5T = 0.0001 (5T = 0.001 (5T = 0.005
0.0001 16.4 25.8 75.6
0.001 15.8 25.8 75.6
0.005 8.0 7.0 71.0
Under t he condit ion t hat pr actically every distinct tr ansit ion in the t rain-
ing set is known ( (58 = 0.0001), t he retrieval accur acy for 500 newly generated
st rings was found to vary from 91 percent wit h (5T = 0.001 to 98.6 percent
wit h (5T = 0.008, a relatively relaxed matching requirement .
The st orage efficiency is sugges ted by t he fact t hat for t he total 4501
symbol transitions in t he 500 st ored sequences, t he required number of V
regist ers ranged from 872 wit h (58 = 0.0001 to 476 wit h (58 = 0.005.
To st udy t he ability of t he TSP t o reject false sent ences, we t rained t he
TSP as before, st oring embedded Reber sentences cont aining t he T-T and
Pi P deterministi c corr elat ion bet ween t he second and penulti mat e symbols.
For t he GSR or recognition st age, however, t he embedded Reber grammar
sour ce was reversed in the penult imate symbol only, guaranteeing t hat none
of t he test sequences had been seen completely during t raining but t hat all
sequences were grammati cally identical up to the penult imate symbol.
Upon performing t he experiment with t his modification, all recognized
sentences (false knowns) corres ponded to false posit ives as indicated in Ta-
ble 1.
When (5T is small , the narrow acceptance window dur ing retrieval easily
blocks acceptance of most nonst ored cases , lowering t he false-posit ive rate
to t he 8 t o 16 percent range. Wi t h a wider acceptance window, however ,
t he false-positi ve rat e soars . These result s suggest a good compromise in t he
vicinity of (5. = (5T = 0.001, alt hough t his choice would int eract wit h t he free
const ants used in t he reacti on-diffusion equations .
4. 6 Demonstration of t ole ran ce to warp
In t he preceding tes ts , one step per symbol was used duri ng both st orage and
testing. This condit ion permits t he symbol concent rations appearing during
GSR st eps to repro duce exact ly t he same sequence of symbol concent rations,
G(t ), as t he net work experienced during st orage, assuming t he tolerances (58
and (5T are equa l. The desir abi lity of tolerat ing int ersymb ol t ime var iat ions
or war p (noted as quality 2 in sect ion 2) is obvious, particularl y if t he original
sour ce of t he symbols is a biological one wit h it s typical random prop ert y.
Consider t he effect s of warp in the ReDi TSP syst em. Let t he spacing
between symbol s., and Sn +l be !:::. t duri ng t raining. Next consider t he int ro-
ducti on of 50 percent warp dur ing recognit ion by GSR. Aft er mat ching Sn ,
t he test for Sn+l occur s at 1.5!:::. t seconds later. The addit ional 0.5!:::.t seconds
A TSP Bas ed on the Bi ological Reaction-diffusion Process
Table 2: Percentage accuracy in predicting t he penultimat e symbol
in an embedded Reber sentence when warp is int roduced before the
critical symbol. All sentences were stored wit h a s = 0.0001. aT is
t he t est (or retr ieval) tolerance. The term 100 % warp means t hat t he
t ime between the test symbol and its predecessor is twice as long as
during st orage. When retrieval is at tempted at wider tolerances, for
example, aT :::: 0.005, the accuracy is st ill excellent up to 200% warp.
aT
No Warp 100% Warp 200% Warp
0.001 91.0 69.6 47.8
0.003 92.8 77.6 63.6
0.005 96.4 95.0 93.2
0.008 98.6 98.8 99.0
0.01 99.6 99.6 99.8
321
will result in a difference of ,0. G, between t he current Gand t he value stored
for t he transit ion s.; ---+ sn+1 in t he same cont ext . The magni t ude of t he error
,0.G will depend on t he free paramet ers in t he reacti on-di ffusion equat ions,
whi ch are not linear. But II ,0. GII will increase, in general, wit h warp. When
II,0.GII > ar , t he warp will have exceeded the abilit y of t he TSP t o recognize
the tr ansit ion Sn ---+ S n+l as valid.
An exp eriment was conduct ed to measure t he relati onship of the error to
warp. The sequences generated by the embedded Reber grammar generat or
were st ored using one ste p per symbol." Duri ng retrieval, however , two or
three t imes as many symbol steps were applied for t he t ransit ion to the
penultimate symbol, effect ing a warp of 100 to 200 percent between t he
(n - 2)th and (n - l )th symbols. The correct or incorr ect recogniti on (known
or unknown) of t he (n - l )th symbol was the err or criterion. The ret rieval
err or rate was measur ed as a fun ct ion of the tolerance, a T (see Table 2). As
aT increases, we are permitting an increasing tolerance for acceptance of the
comparison of a stored sequence with t he external sequence.
The results show t he err or to be rat her high when t he retrieval tol erance
is t ight (aT S 0.001). Relaxat ion of aT lar gely overcomes t he warp bu t
decreases depth and t he corr esponding resolution of sentence distincti ons.
Somewhat greater warp tolerance might be obtained wit h t he use of slower
learni ng rather than one-shot learning. This is one of many trade-offs t hat
can be juggled by t he choice of storage and test tolerances.
4.7 Sequentially addressed sequence memory
We defined t he SASM mode in sect ion 2 as t he case in which t he stored
sequences are relati vely long and t he content address is sequent ially pr e-
sented. The responsive sequence(s) ar e then retrieved. Addressing follows
6Four int egrations per symbol step were used. This permitted each symbol's effect to
reach anywhere in a 5 x 5 array of cells.
322 Sylvi an R. Ray and Hillal Kargupta
the guided sequence ret rieval (GSR) algorit hm, leading t he TSP t o recapit-
ulat e the training steps and set up the symbol concent rat ions that occurred
during training. Subsequent ret rieval of the responsive sequence is performed
by free sequence retrieval (FSR) .
In FSR, there is no exte rnal symbol input. At the end of GSR (the ad-
dr essing operation), t he symbol concent rat ion, G(x, y), has been established
and the pro cessor t hen follows FSR, which is a sequence of cell activat ions
det ermined solely by global-maximum activation (global WTA) , analogous
to a falling domino pat t ern.
The FSR algorit hm follows. Its basic simp licity is complicate d by t he
rat her st ringent demands we will place on it by choice of t ask, which will be
discussed short ly.
Free Sequence Retrieval Algorithm
1. Find the winning cell (most highl y act ivate d globally), C, f C,
corresponding to
N
'v' C : min(ll(V
k
- G(x, y)ll ) = MIN
k=l
where k extends over all V regist ers in all cells and MIN is
t he global minimum.
2. Tentative path A: Emit a uni t of symbol SCi at (x , y) and
Diffuse one symbol step. Find the most highly activated cell,
C
j a
, and the corresponding minimum act ivat ion, MIN
A
.
3. Tentat ive path B: Diffuse one ste p (wit hout an Emit) . Find
t he most highly act ivate d cell, C
j b
, and the corres ponding
minimum act ivat ion, MINB .
4. Select path A if MIN
A
< MIN
B
and MIN
A
< a T.
Otherwise, select path B if MIN
B
< MIN
A
and MIN
B
<
ar. Otherwise, t erminat e with no path.
5. Actualize the steps of the winning path, that is:
Designat e C
j a
or C
j b
as t he AC. If path A, act ualize t he
Emit + Diffuse one step; else act ualize only the Diffuse st ep.
6. Return t o ste p 1.
4.7.1 Selection of a data t ype
There are several questions which deserve to be dist inguished because t hey
pr esent different prob lems.
Addressing: Can address ing utilize eit her initi al or int ernal subsequences?
Sequence complexit y: Is symbol duration st ored and can it be retrieved?
Does the pr esence of subst rings that are common within st ored st rings
cause any difficult y for retrieval ?
A TSP Based on the Biological Reaction-diffusion Process
THE ERIE CANAL
Figure 9: Example of musical melody encoding. A superscript
represents time lengt h in eight h not es. p2p2 is an example of a
"reat t acked" not e. One eighth not e corresponds to a fixed number
of int egration st eps (usually four) of the react ion-diffusion equat ions.
323
To st udy all of these questions, we looked for a data type for which all cases
can ap pear . Such a dat a type is t he musical score of simple melodi es.
First, a system for encoding a single melody of a music score was devised
(see Fi gure 9). The music uni t t ime was chosen to be an eight h not e of t he
react ion-diffusion equat ions . For each eight h note, four int egrati on st eps of
t he reacti on-diffusion equat ions were performed. Four steps ensures t hat an
emit t ed pulse of symbol will t ravel to every locati on in t he TSP st ruct ure
so t hat any next symbol can link wit h it s predecessor. The appearance of a
base symbol (e.g., C, D, F#) was coded as an act ivat ion of t he single cell of
the corr esponding symbol (e.g., A, B, C# ). Relative t ime length in eight h
notes is denot ed by a superscript ."
To distinguish held not es from reattacked not es not ati onall y, we will writ e
C
2C
to represent a C quart ernot e followed by a C eight hnote , and C
3
rep-
resent s a dot t ed quart ernot e or , more pr ecisely, a C t one held for 3 uni t s of
t ime.
At t he storage/training algorit hm, level, t he two sequences C
2C
and C
3
ar e int erpret ed different ly. The former means "Emit C and Diffuse two st eps ,
then Emit C and Diffuse one ste p," whereas t he lat t er means "Emit C and
Diffuse t hree ste ps ." "Diffuse one ste p" corr esponds to four int egrations of
the reaction-diffusion equat ions in t he following experiment s.
An example of a music score t o not ati on mapping is shown in Figure 9.
4.8 Experiments with SASM
Test 1: Distinguishing Sk from sss s (k times)
For t he first t est t he following st rings were stored using O's = 0.0001. The
par amet ers used were D
g
= 0.4, Dr = 0.3, 0: = 0.3, (3 = 0.3, E = 0. 2, an d
"y = 0. 1.
ABC D
E BCD
G BCD
7For a single eighth note, superscript 1 defaults to blank.
324 Sylvian R. Ray and Hillal Kargupta
The sequences were addressed (by GSR) up to the I , and t he remainder of
t he sequences were ret rieved by t he FSR algorit hm. The object ive was to
t est t he ability of t he TSP t o distinguish t he F F F case from t he F
3
case
based on the init ial symbol of the address, E or G.
The retrieved st rings were exact ly as st ored. The sequence of G vect ors
experienced dur ing ret rieval was exactly t he same as occurred during storage
since IJ
s
was small enough to give a depth of mor e t han 4, t he dist ance from
t he last unique symbol to t he point where t he sequences diverge.
In t his case, not only does t he difference among t he initi al four symbols of
each sequence pr edict the fifth symbol correctly, but t he more difficult task
of dist inguishing F F F from F
3
is achieved.
Not e t hat there is no problem due to t he common (BCD) subsequences
here or, in general, as long as t he effect of t he first symbol is not confounded
by t he depth being too small (i.e., by IJ
s
being t oo small) .
Test 2: Variable length symbols in the addr ess
Suppose t he same number of appearances of a symbol occur s in two ad-
dresses, but t he t ime lengths of symbol appearance are different . Can t he
TSP avoid confusion? To t est t his, we store the following two sequences.
AB C F F- 2 FIE D C
AB E F F F FIG F E
Again, using IJ
s
= 0.0001, which is small enough to ensure t hat t he context
of each t ransition is uniquely repr esent ed, t here was no problem at all in
ret rieving t he corr ect sequence following t he I by FSR. All par amet ers were
as in t est 1 except t hat D
g
= 0.3 pr oved to be a bet t er choice.
Test 3: Sequences wit h long common subsequences
Fin ally, we stored sequences having common subsequences to test t he
ability of t he TSP to avoid confounding t he sequence during retr ieval. The
stored sequences, with the address sect ions shown, are
F A-4 I B D E F G B C D-S
F- 2 D IE F G A-3 B- 4
F-3 C IDE F G A- 2 B-3
B- 2 C-3 E- 2 D E F C-3 B-2 A C
B-2 C-3 E- 2 D E C-4 B-2 C A
B-3 C-3 E- 2 D E C-2 E-4
The common subsequences were in the free ret rieved section in some
cases and in t he address section in ot her cases . Using the same paramet ers
as before, with t he except ion t hat D
g
= 0.2, all sequences could be retrieved
exac t ly-meaning both t he symbols (tones) and t heir duration.
4.8.1 Comments on t he FSR algorithm
We can now appropriately discuss t he complication in t he FSR algorit hm
involving the need t o t est t wo possible actions at each st ep (see steps 2 and 3
A TSP Based on the Biological Reaction-diffusion Process 325
in t he FSR algorit hm, section 4.7). If one permit ted only sent ences having a
single symbol per st ep, t he simple rule of the emission of a uni t of a symbol
on each t ransit ion would be adequate. The need for the t est arises from t he
fact t hat we want to dist inguish cases of "held" not es versus "reatt acked"
not es. For example, in t he subsequences . . . BC
4
B . . . and . .. BCC
2C
B . .. ,
each step t hrough t he region of Cs requires testing whet her t he act ive cell
supports a path corresponding to (Emit + Diffuse) = "reat t acked not e" or
a path corresponding to (Diffuse only) = "held note." If t he sequences were
restr icted to only an "at tacked note" for each symbol transit ion, t he test
of alternat ive possibilit ies seen in t he FSR algorit hm would be unn ecessary.
But wit h t he allowance of successive appearances of t he same symbol eit her
held or reat t acked, t he way to distinguish t he pr eferr ed pat h is by t est ing t he
two possibilit ies corres ponding to path A and path B in t he FSR algorit hm.
Also not e t hat t he TSP does not support addressing t ha t begins at an
internal subsequence of a stored string. The reason for t his limi t ation is th at
an approxima te value of t he G vector at t he int ernal symbol where addressing
begins would have to be known in order to conti nue t he FSR correctly. For
addressing wit h an ini ti al substring, t he addressing always begins wit h G =
and the GSR develops t he complex distribut ion of symbol concent rat ions .
5. Discussion
We have proposed and st udied a temporal sequence processing system t hat
uses the reaction-diffusion pr ocess to provide interconnect ion of all cells and
convey memory of past events .
The cells are provided wit h a mult iplicity of input s that are capable of
act ivat ion due to t he symbol concent ra tion at t he locat ion of t he cell, pro-
ducing a cell like a multi -inpu t radial basis function cell. React ion-diffusion
provides for controlled growth and decay processes, rather t han just decay
and diffusion, for t he concent rations, which adds another dimension of possi-
bilit ies to the sequence representati ons contained in t he TSP and is a unique
feature. Ot her designs having similar object ives [10, 11] limit t hemselves to
monot oni c decay of memory of past event s.
The expe rimental st udy of t he ReDi TSP was keyed to five qualit ies,
which are, br iefly, depth , flexibility of equivalence class repr esentat ion, warp
t olerance, minimum stor age, and content addressability. Various combina-
tions of t hese qualiti es are pr eferred for various applicat ions.
The present system permi t s any dept h desired but at an increasing cost
in storage capacit y for increasing depth. A unique feature of the system is
t hat it allows flexibl e class equivalence at retrieval t ime. Once t he sequences
are stored with a specific degree of uniqueness of t he tr ansiti ons, they may
be retrieved (or recognized) using t he same or a lesser degr ee of stringency
in symbol mat ching.
Storage efficiency, in relative terms, is controllable by a single par ame-
ter , o"s , which sets t he radius in symbol concent rat ion space within which
sequence transit ions are recogni zed and stored as unique. Thus, storage
326 Sylvian R. Ray and Hillol Kargupta
efficiency and it s inverse, depth, are easily cont rolled by select ion of t he pa-
ramet er O"s.
The system shows a moderat e amount of warp tolerance, but t his is
achieved at t he cost of accuracy in identifying t he stored sequence.
Two principal t asks were examined for t he purpose of evaluat ion of t he
system: embedded sequence recognit ion (ESR) and sequent ially addressed
sequent ial memory (SASM). InESR, st ored short sequences ar e examined for
t heir occurrence in a longer argument st ring. Bot h tasks are accomplished by
combinations of t he more fund ament al algorit hms, guided and free sequence
ret rieval (GSR and FSR).
For t he ESR problem, t he external sequence guides t he inquiry (address-
ing), which is, operationally, guided sequence retrieval. This algorit hm was
tested using an embedded Reber grammar to generate short sequences for
storage. Other sequences were applied as arguments t hat , alt hough not nec-
essarily physically longer , provided an equivalent test to t he case of an ex-
tensive argument st ring. Successful events were t hose t hat correctly reached
and pr edict ed t he penul ti mat e symbol. This t est was fully successful, reach-
ing accur acies up to 98 percent for parti cular values of o, and O"T . A t est of
false-positi ve responses showed that the accuracies at t ributed to true-positi ve
cases were largely valid.
For t he SASM problem, t he external sequences are short and the int ernal
sequences are long. The external sequence is used as an address, applied
by t he GSR algorit hm, and t he remainder of t he int ernally stored st ring is
retrieved by FSR in a method analogous to following t he lowest energy path
from t he point where t he address ends. We confined ourselves to addressing
from t he initial subsequences of the stored sequences only; addressing begin-
ning int ernal to th e stored sequences is not possible, in general, with t he
TSP.
8
The TSP has no problem at all performing t he SASM problem for ordi-
nar y complex sequences that are stored and retrieved wit h one symbol per
t ime step. In order to demonstr at e the greater capabilit ies of t he system, we
int roduced and st ored simple melody-like sequences requi ring that bot h t he
symbols and t heir t ime duration be correctly retr ieved. This experiment was
also fully successful, alt hough it required an unwant ed complicat ion in t he
FSR algorit hm in t he form of a test-and-choose pr ocedure to correctly iden-
t ify the "held" versus t he "reat t acked" case when t he same symbol recur s.
Comp aring t he TSP to t he met hod pr esent ed in [11], t heir approach ap-
pears to be mor e tolerant t o warp t han ours. Their approach requires pre-
liminary knowledge of the necessary depth before commencing st orage of
sequences. The feature of t he TSP method t hat allows adjust ment of recog-
nition tolerance during t he recognit ion phase is not supported, as far as we
can det ermine, in t he meth od in [11]. The import ant pr operty of "chunk-
ing," however, is one we have not yet addressed, alt hough we do not see any
inherent difficulty in extending the TSP to perform chunking.
8 A modified system t hat permits addressing from intern al sequences has been st udied
by us and will be present ed elsewhere.
A TSP Based on the Biological Reaction-diffusion Process
R eferences
327
[1] Turing, Alan , "The Chemical Basis of Mor phogenesis, " Phil . Trans. of the
Royal Soc. of London, Ser . B, 237 (1952) 37-72.
[2] Ray, S.R. & Gyaw , T.A. , "Universal Multi channel Signal Interpretation Using
Connectionist Architect ure," in Proceedings of Artificial Neura l Networks in
Engineering Conj. , V.4, (ASME Press , New York , 1994).
[3] Cleeremans, Axel, Mechani sms of Implicit Learning (MIT Press, Cambridge,
MA,1993) .
[4] Mozer , Michael C. , "Neur al Net Archit ectures for Temporal Sequence Pro-
cessing ," in Predicting the Future and Understanding the Past, edited by A.
Weigend & N. Gers henfeld (Addison-Wesley, Redwoo d City, CA, 1993).
[5] Elman, Jeffrey, "Finding St ructure in Ti me," Cognitive Sci ence, 14 (1990)
179- 211.
[6] Jordan, M.L, "At t ract or Dynamics and Parallelism in a Connectionist Se-
quential Machine," in Proceedings of 8th Annual Conference of the Cognitive
Scienc e Society, 1987, 531-546.
[7] Cummings , F. , "Representation of Temporal Patterns in Recurrent Neural
Networks," in Proceedings of the 15th Annual Confe rence of the Cognitive
Science Society, 1993, 377-382.
[8] Fahl mann, S.E., "The Recur rent Cascade-correlation Architecture," in
Advances in Neural Inf ormation Processing Systems; NIPS-3, Proceedin gs
of the 1990 Conference, held at Denver on November 26-29, 1990, edit ed by
Lippman n, Moody, and Tour et zky (Morgan Kaufmann, Palo Alto , CA, 1990).
[9] deVri es, B. and Principe , J.C., "The Gamma Model-A New Neural Net Model
for Temporal Processing," Neural Networks, 5 (1992) 565-576.
[10] Bodenhausen, Ulrich and Waibe l, Alex, "Learn ing t he Ar chit ecture of Neural
Networks for Speech Recognit ion," IEEE Proceedings of the I CASSP, 1 (1991)
117-124.
[11] Wang, DeLiang and Arbib, Michael, "Timing and Chunking in Processing
Temporal Order ," IEEE Transcrip ts on SMC, 23 (1993) 993- 1009.
[12] Meinhar dt, Hans, Models of Biological Pattern Forma tion (Academic Press,
1982).
[13] Bohn, H. , Interkalare Regeneration und segmentale Gradient en bei den Ex-
t remitaten von Leucophaea-Larven (Bl attaria) . III . Die Herk unft des in-
terkalaren Regenerats. Wilhelm Roux Arch. , 167 (1971) 209-22l.
[14] Fr ench, V., "Leg Regeneration in t he Cockroach, Blattella germanica L Re-
generation from a congruent t ibial graft/host juncti on," J. Embryol. expo
Morph., 179, (1976) 57- 76.