IEE TRANSACTIONS ON CIRCLTTS AND SYSTEMS: FUNDAMENTAL THEORY AND APPLICATIONS, VOL. 4, NO 0, OCTOBER 1995 na A Computer-Assisted Investigation of a 2-D Array of Chua’s Circuits Fatih Kavaslar and Cuneyt Gizelig Abstract—This paper presents simulation results on a 2-D array of coupled Chua’s circuits called a chaotie Cellular Neural Network (CNN) in [1], but here will be called an f(r) ~ coupled chaotic CNN in order to distinguish i rom other recently proposed chaotic CNN architectures (2]-[4]. Each isolated cell with unity self-feedback in the network is a Chua's cireult and is ‘connected only to its nearest neighbors defined by a metric and ‘a neighborhood size [I] [5]. The network is designed with a 2-D torus like connection topology having eyelic boundary conditions that play an important role in complete phase syncronization. It 's observed in our computer simulations that i) depending on the choice of the intracell parameters and the connection weights, the cells of the network appear to be operating in a double-sroll (Chua's attractor, in spiral Chua's attractors, in stable equilibria, ina period-1, a period-2, a large limit-eyce, or ina large chaotic regime, i) complete phase synchronization in the network with all cells operating in the double scroll regime can be obtained by a sultable choice of the intracell parameters and the feedback connection weights, ill) there I a set of the intracell parameters ‘and connection weights resulting in a chaotic regime such that ‘each cell depending on its constant external input falls into one of. the three attractors; namely, the double-serol, P* spiral, or P” spiral Chua’s attractor. AS a new phenomenon, a close relation between phase synchronization setling-ime and input pattern is ‘observed that offers new potentials of Chua's cireuit arrays for pattern recognition applications. 1. INTRODUCTION ETWORKS OF coupled chaotic subsystems have been investigated as a new paradigm for high-dimensional ‘chaos, for biological neural networks, as well as for infor- mation and signal processing {6]-(8]. Among such networks, arrays of Chua's circuits (1]-[4} are gaining particular impor- tance under the motivation of a great deal of investigations on, its dynamics and hardware implementations [9}-{12]. In recent studies [2]+{4), synchronization and hyperchaos phenomena have been observed in various arrays of Chua's circuits. This, paper reports several qualitatively distinct dynamical regimes and complete phase synchronization phenomena observed in, computer simulations of a 2-D array of f(x) — & coupled Chua's circuits. This network of coupled Chua's circuits is, indeed a 2-D version of the chaotic CNN reported in [1] and constitutes a special case of the generalized CNN's introduced in [5]. It reduces to the CNN [13] upon a modification of some intracell parameters and slopes of the segments in the piecewise-linear output function. Manuscript received January 15,1995; revised April 26, 1995. This paper vas recommended by Guest ator L.O. Chua “The author ae with the Faulty of Electical-Elecwonics Engineering ‘stant Technical Univerty, Maslak 80626, Tukey. TEEE Log Number 9414468, ‘The Key features of the CNN considered in this paper are: i) the possibility of processing constant and also time- varying 2-D signals fed via external inputs, i) the possession of a CNN-like connection topology with a space-invariant connection weight pattern, and ili) the possession of an f(z) — type of coupling that ensures the bounded-input bounded- ‘output stability of the whole network, which might have inhibitory connections or excitatory connections. ‘The 2-D array of f(e) ~ # coupled Chua's circuits con- sidered in this paper differs from other known Chua's circuit arrays (2), [3], [14], [15] in the following respects: i) The models in (2), {3], [14] are 1-D arrays of Chua's circuits; but the chaotic CNN considered in this paper is a 2-D array of (Chua's circuits. ii) In the 2-D array of Chua's circuits used in [15], the individual cells are designed not to be operating in the double-scroll regime even when they are isolated. In contrast, each isolated cell in our chaotic CNN operates in the double-scroll regime. iii) As a direct consequence of the ‘unbounded 3-segment piecewise-linear nonlinearity used in the ‘models of [2}-I4}, [14], [15], only passive resistive couplings, i.e, only excitatory connections! are allowed in these models. ‘The chaotic CNN can have inhibitory connections as well as excitatory connections. iv) The 2-D array model in [4] uses the canonical Chua's circuits as the cells and uses both =~ & ‘and ~ 2 linear couplings. v) The coupling in our chaotic CNN is nonlinear as opposed to other known Chua's circuit arrays (2]-(4], [14], [15]. The output f(z) of each cell in the cchaotic CNN is fed to the neighboring cells resulting in a forcing term in the first equation of (1) describing each cell, ‘Therefore, the coupling in the CNN considered in this paper is called an f(z) ~ # coupling. ‘The state equations describing each cell of the 2-D chaotic CNN considered in this paper is defined by the following dimensionless form: aygsae|-Fenstust DD ana Seis ga) sie(toay + bse wsnsl +7 Oy stetouy thy Bs Ms ts @ fy Bis @ "ne conductance ofa passive liner resistor f postive, hence the cute, ‘passing trough this essor combate a postive vale tothe ttl input ofa ‘osesSoaptc cel ifthe utp of the assciated presynaptic cells pastivem [BEE TRANSACTIONS ON CIRCUITS AND SYSTEMS se) Fig 1. The characteristics ofthe 7-egmentpicewite-lner function f(). ‘The outermost zeroslope segments coincide withthe hoizotl axis where Sle) = 5 -(rmo—m:) (le+ | ~ | —1) 4+ 5m ma): (le + Bil [2 ~ Ea) + 5 +ma+(le-+ Bal - [2 ~ Bal). ® Here, i € {0;1,2,-...M—1} and j € (0,1,2,-.-,N=1) for an M x N network. In order to obtain cyclic boundary conditions, the network is designed as a 2-D torus-like aay. ‘As observed from our simulations, cyclic boundary conditions prevent the disturbing effect of boundary conditions on the complete phase synchronization. The external inputs 1 (0)’s might be time-varying. The constant threshold I sidentcal for all cells. The ag and by, coefficients define a space-invariant feedback connection weight patern and a space-invariant input connection weight pattem, respectively. In the sequel, these coefficients will be specified in the standard matrix forms A and called the feedback and input templates, respectively A mae bo baa ) B boo bor bo bia ‘The output function f(-) defined in (4) is a piccewise- linear function different from the original characteristic of the ‘Chua's diode. As can be seen from Fig. 1, the function f(-) is bounded and lies eventually on the z-axis. The reason for choosing this bounded function is to ensure the bounded-input ounded-output stability of the overall network for inhibitory connections as well as excitatory connections [1]. On the other hand, such a choice of the output function results in ‘an eventually passive Chua’s diode for an isolated cell [1]. ‘The outermost zero-slope segments and the negative-slope segments have no effect on all attractors observed from a three-segment characteristic if Ey is chosen to be less than three. These segments add a large limit-cycle and a large chaotic attractor (see Section IB) into the dynamical behavior : FUNDAMENTAL THEORY AND APPLICATIONS, VOL, 2, NO. 10, OCTORER 1995 repertoire while ensuring a bounded-input bounded-output stability, ‘The dimensionless equations in (1)-(4) can be obtained as a ‘two-dimensional special case from the equations defining an n- dimensional chaotic CNN [1] by setting the time delays to zero while choosing the neighborhood size equal to one and also ‘making the following changes of variables and parameters: He Es a ¢ HE a = "Bs = Bs a= = Lt Bs Me = thi = Ba = Bs 2, However, he notation t, E's, and m,'s are used inthis paper instead ofr, B's, and E's, respectively “Throyghot this pape, the slopes are chosen a my = 32, mp = —1; the breakpoints are located at E, = 20 and E2 = 381, For these slopes and breakpoints together with § = 2 and'a, = 1, an isolated cell of our chaotic CNN’ when restricted to the region specified by je, | < Ey becomes identical to the original Chua's circuit, Here, an “isolated” cell is defined as a cell that has no external input and has no connection with other cells. Moreover, inthis paper, the tem “uncoupled” cell will be used for a cell that has no connection with other cells but has external inputs; and the term “fre cell willbe used for a cell that has no extemal input, and itis ‘not excited by any other cell, but its output is fed 1 some cells {Al experiments reported in the paper were done for I a =9, 6 = 2, The dynamics of the network are investigated for several choives of connection weights, external inputs, and intracell parameter . In all of our simulations, the Forward-Euler method with the step size of 0,001 was used for integrating the differential equations (1)-( Several atractors obtained from an uncoupled cell excited by constant, sinusoidal sources and also by the output of free cell operating in the doublescroll regime are presented in Section Tl An experiment showing a coherent dynamical behavior of all cells operating in the double-scroll regime is also presented in Section Tl, Simulation results showing a complete phase synchronization and its application to the recognition of lines in images are described in Section I UL. DEPENDENCE OF DYNAMICS ON CONNECTION WEIGHTS AND INPUTS (Our analysis in this section is twofold: one is the analysis of the dynamics of an uncoupled cell as a function ofthe intracell parameters and inputs; second is the analysis of the overall network based on the connection weights and inputs. The ultimate goal of our analysis is to design a network of identical cells such that the dynamics of a single cell are controlled by an intracell parameter, for instance f, and the dynamics of the overall network is controlled by the connection weights A. Double-Seroll Chua's Attractor as a Coherent Mode We first investigate if the double-scroll regime can be a coherent mode of operation for coupled cells each of which is individually designed to be operating in a double-scroll regime. Such a coherent dynamical behavior of all cells in the double-scroll regime was observed in several experiments with different choices of the feedback connection weights and network sizes. One of these experiments was realized for anKAVASLAR AND GUZELIS: ACOMPUTER-ASSISTED INVESTIGATION OF A 2-D ARRAY OF CHUA'S CCL! (2.4102, 0.41034) .4102,0.41034) (2.4102,2.4102) x33 CRIs waa) o (2.3256 0.41509) 42.4006,0.43794) (2.3256,2.4006) fi fi x3. 2.2057, 0.42660) a e388 84TH a3 CERO E37 oi o © ao Fig. 2. The tajecory plots (Lissjous furs) foram 8 8 chaotic CNN operating in 2 coherent doble-seol| mode (8) 2-1 tegectory plot for ell C(1,1) under uniform iia condos. ¢D) 2-y wajectory plot for cell C(3,3) under nifr ina tno (c) x acta plo for he pir of eels C(I, 1) and C(3, 3} under uniform ital conditions. ()2—y trajectory pot foveal C(21)wnder noun inal conditions (e) 2 uajectry pot for cll and 3.3) under sonniform inl Condens (0 - tjectry pls for the pair of eels C(L-1) and (3,3) under noni iil endtons. 8 x 8 network with # = 14.28. The templates used are as, 13, follows we (ine o FP Saran no oo onal lig Lint Cyces wound Pend F Ar= {0.00 1 0.001 0.001 0.001 0.001 1 0 0 0 Bi=]0 0 0 © as} 000 a A set of trajectory plots (Lissajous figures) obtained with the a] Periods Lint Cycle wound Pfand initial conditions 2, 5(0) = yij(0) = 2i3(0) = 0.1 (identical for all cells) is given in Fig. 2(a)-(D. As can be seen from a8) nb Basti the z-y trajectory plots in Fig. 2(a) and (b), the double-seroll, 3 regime of the isolated cells with the parameters @ = 14.28 {7 28 «6 6 vow and a,» = 1 is not destroyed in a network of weakly- coupled cells. Foran identical choice of the inital conditions, not only mode coherence but also the phase coherence was observed. The phase synchronization observed for the pair of cells O(1,1) and C(3,8) is illustrated by the 2-2 trajectory plot in Fig. 2c). For the above parameters, all cells operate inthe same mode and with the same phase, However, for the considered 8 x 8 network, complete phase synchronization is seen to be sensitive to initial conditions so that it may be destroyed by choosing initial conditions which differ from Fig. 3. A tough 3 — ane bifurcation diagram, fone cell to another. Fig. 2(d)-(f) shows the destruction of the phase coherence between cells C(1,1) and C(3,3) and the preservation of the mode coherency under nonuniform initial conditions. A 3 x 3 network exhibiting complete phase synchronization that is persistent to variations in the initial conditions and extemal constant inputs will be reported in Section Ima IEEE TRANSACTIONS ON CIRCUTTS AND SYSTEMS.—t: FUNDAMENTAL THEORY AND APPLICATIONS, VOL. 42, NO. 10, OCTOBER 195 (2.2487 0.40401) (@.900ai,-0.34037) (60502-2295) = CEns3, 0-005) « © © (4.8092,0.24074) COTTas, -0. 79648) = Ta OSaaSey ® (27.051 26.659) (6.0196 1.3698) PTO FET = 8s = aT x ® » o Fig. 4. ‘Typical y trajectory plots obsinod for several choices of and do. in an isolated cell (a) A trajectory converging ‘ovard the P* equitrium. () An elite limiteycle suroundig the P* equilitriam, (e) A double-scoll Chu's atractr (8) A P* spiral Chua's atactor. (@) A'perod2 lmatcyele around the P~ equlbeum. () A peiodt limi-eycle around the P+ equim. (g) A trajectory converging toward a lage limieyele. ¢h) The newly observed lage chai atactor G) A, trajectory converging toward the P* equilinm, B, Attracts in an Isolated Cell ‘A set of experiments is done for several choices of the Investigation of the coherent modes in the overall network Paameters and ay. The other intracell parameters are as requires the knowledge of the dynamics of an individual &!¥e" in Section T, and the templates are cell. An isolated cell with unity self-feedback, ie. ay = 1 ao 0 000 investigated extensively, thus, there is no need to consider the 0 So 000 unity self-feedback case here. However, an isolated cell with 40.0 # 1 differs from the original Chua's circuit and needs to ‘The bifurcation diagram in Fig. 3 roughly specifies the re- be investigated further. sions of parameters for which an atractor exists, All atractors[KAVASLAR AND GOZELIS: A COMPUTER ASSISTED INVESTIGATION OF & 2-D ARRAY OF CHUA'S CIRCUTTS (2.7065 0.58795; (26.326 ,24.377 = (72,046, -0. 6144ay * @ Fig. 5. The <-¥ trajectory plots for an uncoupled cell diven by ‘obtsine fr 15 00 iteration steps starting from tential inital conditions (0) Weight bye is OATS. (b) The slfnpot conection weight Boy. is 0.078 reported in this diagram have been observed for a set of initial conditions chosen randomly in a bounded set enclosing the origin. The large limit-cycle illustrated in Fig. 4) survives for initial conditions chosen outside of this region while all other attractors disappear. Some periodic windows in the double- seroll parameter region were observed but not shown in the diagram, A set of the observed attractors is displayed in Fig. 4. Al of the trajectories were obtained for 40000 iterations starting from the initial conditions 2,3 = yay = 245 = Ol and plotted after discarding some initial transients. It is interesting to note that the P+, P=, and P° stable equilibria the P'* and P= Chua's spiral attractors the period: rlimit-eycles, the double-scroll Chua’s attractor. and the large limit-eycle (see Fig. 4(a)~(g) and (j)) observed in an isolated cell with ao¢ # 1 are also included in the dynamics of the original Chua's circuit. On the other hand, an isolated cell, with unity self-feedback possesses all dynamics of the Chua's cireuit since they are identical. Furthermore, these dynamics, can be produced by an isolated cell not only for unity self- feedback but also for a wide range of self-feedback weights AA new attractor not observed in the original Chua's circuit is the large chaotic attractor shown in Fig. 4(h). This attractor is, inherently due to the negative-slope segments In the light of our study on the 2 — do.e bifurcation of the dynamics, one could search for coherent modes in a network of coupled cells all operating in one of the attractors mentioned above, The coherent double-scroll Chua’s attractor rode described in Section ILA and the coherent large limit cycle mode were observed in simulations of several types of couplings. Our investigation of coherent modes was basically ad hoc. However, a necessary condition for mode and phase coherency will be given in Section II, C. Excitation by Free Cells ‘The dynamics of an uncoupled cell driven by a free cell ‘operating in a double-scroll regime are of particular interest not only from the point of view of the network dynamics under ‘excitation but also from the coherent double-scroll mode under rho excitations. In a coherent double-scroll mode, each cell (26.325, -24, 396) * o ting inthe double-scoll egime, The pots wore (0) = 2{0) = 01.) The self ipat connection 15 boo | Large Limit Cycle : ‘Attractors as o Double Scroll Chua's Attractor 05 Attractor A ‘Large Limit Cycle — 7 . 107 10 10 10 Fig. 6. A rough fo — Baw bifurcation diagram. Cli, 3) is excited by a signal gi) that is a weighted sum of the f(2;5(E)) waveforms associated with the neighboring, cells that are all operating in a double-sroll regime. Therefore, an analysis ofthe cells excited by a free cell provides some insight into the coherent dynamics. Tn an experiment done for the following templates and (9 = 14.28, it was observed that small input connection weights do not destroy the double-scroll regime as seen in Fig. 5(a). On the other hand, large input connection weights force the cells to be operated in the large limit-cycle as illustrated by Fig. 5(b). 0 0 0 0 0 0 As=|0 1 0], By=|0 bo» 0 ® jo 00 o 0 016 (2.0131 0.56657) IPE TRANSACTIONS ON CIRCUITS AND SYSTEMS—1: FUNDAMENTAL THEORY AND APPLICATIONS, VOL 42, NO. 10, OCTOBER 195 (2.3791,0.42291) (72-0206 ,-0. 55426) ¥ (2.5829 ,0.33151) ($0,37082 ,-0 46644) ® © (-2,2072,-0, 30262) * co Fig. 7. The x-y trajectory plots for an uncoupled cell driven by sinusoidal oscillator with unit ample, The plas were obtained for 40000 iteration Seps Saring from identical inital conditons 2(0) = y(0) (09 () fo = 10 and by. O) fe = 1 aod nye FOL. fo = Lad bo For the above templates, an empirical result for an uncou- pled cell to remain in a double-scroll regime when driven by 4 free cell operating in a double-scroll regime was estimated 10 be [B,,o| $ 0.0475. On the other hand, the absolute sum of the feedback connection weights excluding the self-feedback boeing less than 0.0475 can be estimated as a necessary condition for a zero-input network to be in the coherent double scroll mode, Because, in a network of coupled but zero-input cells, the contribution gi,;(#) of the neighboring cells to a cell may be considered as the output of a free cell. Note that the sum 0,008 of the mentioned feedback weights in the experiment of Section ILA is less than 0.0475, which confirms ‘our empirical resul D. Excitation by Sinusoidal Oscillators ‘The dynamics of a single cell under sinusoidal excitations ‘are worth investigating for several reasons. First, monitoring the dynamics helps us to understand the effects of interac- tions between cells. Second, some unknown coherent periodic modes can be discovered during the process. Third, one may ‘obtain useful information about the use of the whole network for processing periodic signals fed via external inputs. 3(0) = Oe (a) fo =F and bye = 0. 3 ‘Several experiments have been performed for the templates in (8) and the attractors observed for several choices of the sinusoidal oscillator frequency f, and the input template coefficient b,,. are summarized in the f, — by,» bifurcation diagram of Fig. 6. Here, /? has been chosen as 14.28, and the input sinusoidal signal has unity amplitude. Four of the observed attractors are displayed in Fig. T(a)-(d). E, Excitation by Constant Inputs This part of the experiments is devoted to the analysis of uncoupled cells under constant stimuli. Here, the used is 14.28, and the templates are as follows: 00 0 fo 0 0 Ar=|o 1 0], By=]o 1 0 ® 000 000 For the above choice of the parameters, a cell under no stimulus is operating in the double-scroll regime as shown in Fig. 8(b). It is also observed that positive (resp. negative) inputs drive the cells into a P+ (sesp., P~) spiral Chua’ attractor. Such a stimuli-dependent switching from the double- seroll Chua's attractor to the spiral Chua’s attractors can beAVASLAR AND GUZELIS: A COMPUTER- ASSISTED INVESTIGATION OF A 2-0 @ ® (2.9359,0.31525) a ("29353-0405 [e098 83688) A © 10339,0.46937) (2.3214,0.61702} 2.2 y0.0 (Ci. 32a3, 0.61023) 52 CEawT “0.35 @ @ Fig. 8, 4.3 x 3 network of uncoupled cols driven by’ constant pus. Pos sey a C2935 -0 1491) a0 © tive its drive the cells into a PY spiral while eave inputs into Psi) Image fed tothe network a external saput. Black coresponds wo the inp level of ~0.125 ‘The oer Bay levels corespond fo—0025, D000, 0.075, and 0.125. (8) The {2 tajestny plot forthe eel C(2. 1). (0) The zy ujectory pot For the eel CQO} th The ay trectory plot for the cll C(0.0). useful for classification purposes. Moreover, a network of uncoupled cells can also be used for encoding three-level images such that each level is associated with one of the three chaotic states; namely, double-scroll, P* and P~ spirals Fig, 8(a) shows a 3-level gray scale image applied to 2 3 x 3 network with the above parameters. The 2~y trajectory plots for some cells are illustrated in Fig. 8(b)-(. Note that the upper-left pixel and the mid-left pixel correspond to the calls C(0,0) and (0,1), respectively; and the other pixels correspond to the remaining cells in the same arrangement. IL, PHASE SYNCHRONIZATION Phase synchronization has been widely investigated in sev eral arrays of coupled chaotic or periodic oscillators as a phenomenon that may play a role in biological information processing and might be useful for information processing in engineering applications. Complete, partial, and turbulent phase synchronizations have been obscrved in some arrays ‘and coding of distinct objects as clusters of synchronized subsystems has been proposed as a possible application. In our experiments, phase synchronizations were observed in the coherent double-scroll mode and in the coherent large limit-cycle mode, We present below simulation results only for a 3 x 3 network exhibiting a persistent complete phase oy trajectory plo forthe cll C(1, 1). (@) The i C(2,2)-(e) The ey eaecory plot forthe cell synchronization in the coherent double scroll mode. The 3 used is 14.28, and the templates are as follows: fo.04 0.04 0.04 00 0 As= 001 0.68 0.01], Bs= [0 0 0}. (10 0.04 0.04 0.04 000 ‘The 2-level 3.x 3 image in Fig. %a) was applied to the network as a set of initial conditions common for all = vs and 2 states variables, ie, 2:,3(0) = wig(0) = 2%, (0) forall i, j. Four different snap shots showing the evolution of the {1-1 trajectory plot belonging to the cells C(0,0) and C(1,1) are illustrated in Fig. 9(b)-(e). As also observed in many other simulations, nonuniform choices ofthe inital conditions across the network do not annihilate the complete phase synchroniza- tion but cause some transition periods before synchronization is achieved. The complete phase synchronization observed is also persistent to variations of the initial conditions from one state variable to another in the same cell A. A Necessary Condition for Phase Coherency ‘The complete phase synchronization in the double-seroll mode reported above has also been observed for many 3 x 3 networks with the same parameters and the feedbackmn IEEE TRANSACTIONS ON CIRCUTTS AND SYSTEMS: FUNDAMENTAL THEORY AND APPLICATIONS, VOL. 42.NO- 10, OCTOBER 1995 oO 1 2 oO 1 2 @ (2.1999 2.2256) (2.2516, 2.2358) 3 8 ia a ® © (2.2219 2.17) (2.2022,2.1942) 2 2 i 3 CRETE) a COTE) aT Fig. 9. Complete pase synchronization obtained bythe Ax ~ Bstemplats. (a) The 3 3 image represents the ital conditions Black and white comespond 1 the zero nO levels respectively. (0) Fist sap sbot showing the rstao unjectry plot for the period hat covers 30000 iterations before the els ate coupled. (c) Second snapshot showing the 10 taestory pt for the period that coves 30000 iteration after the cells are coupled. () Third sap thot showing the 1 ro tyjectoey plot for ‘the pri tht covers 30000 iterations Beginning at 300th seration aftr the cll are coupled) Fourth snp sot showing the {51,s7B0qtajectry plot for the period tha covers 30000 seritions beginning t GDOOOR iteration after te ces ae coupled connection weights satisfying the condition W = 1. Here, Indeed, phase coherency in the double-sroll mode has W is defined as appeared for W's chosen from the set D. Here, D denotes the set ofa. values such that an isolated cll with a self-eedback ant (1D weight chosen from D operates in the double-scrol regime. It ntetoay can be shown by simple analysis thatthe condition of W € D wKAVASLAR AND GUVELIS” A COMPUTER-ASSISTED INVESTIGATION OF 4 2-D ARRAY OF CHUA'S CIRCUTTS 0 1 2 0 1 2 ! (2.2337,0.42139 (2.3657 2.3768) : CARRS AE A eR CRIS =EaTOR) EE (2.2620,2.2367 (2.2400,2.2363) 3 3 a line with slope tn = 1 repesets the external inputs, Black and white corespond to the zero and 0. Ives respectively. (0) A ‘map sho showing the 1.17.9 tajectory plot Tor the peri that coves 30000 leon before the cells are coupled (¢) A ‘Snapshot showing the 11-70 tajetory pot forthe pee hat covers 30000 izations afer the els are couple. Cd) A snap Shot showing the firoo wajectory ple forthe pero that covers 30H iterations besnning at 30000 iteration after the ‘lls are coupled. (6) A stup shot showing the 1, 1=yo. trator plot forthe ered that covers 30000 iterations beginning GOOD eration ater the celle are coupled is, indeed, a necessary condition for a network with zero-input and 2,,(t) = =(t) for all t as required for complete phase to be phase coherent in the coherent double-scroll mode. To synchronization under uniform initial conditions. Now, the do 50, let us assume that 7,(0) = 2), nig ) = ult). dynamics of the overall network are described by the following70 BEE TRANSACTIONS ON CIRCUITS AND SYSTEMS: FUNDAMENTAL THEORY AND APPLICATIONS, VOL. 42, NO. 10, OCTOBER 1995 ¢2.1652,2.2276) <2.2257,2.2861 wat (2.2123, 2.2572) x0 0.5851, -0. 70352) 70 o ® <2.2508,2.2697> <2.2632,2.2568) (0. 47156,-0-046003) 1.0 2. BaST aaa) si-o » @ «2.2632, 0.095024) <2.2452,2.6636) % ¥ 2.2257, 1.8098) o xO xo Fig. 10. (Cominued) (A snap shot showing the 2 0-22, trajectory plot forthe period that covers 200 iteration before the calls are couple. (g) A soap sot showing the y,0~22.. eaecory plot Toe de ped that coves 30000 itecations rhe ves ‘recouped h) A snapshot showing the 0-72 tjetory plo forthe peso tht covers 30000 iterations begining at 30000 iteration afer the celle ae couple. () A snap shot showing te 0-721 taectory pot forthe period that covers 30000 iterations begining at GOOD iteration alr the cells are coupled.) A Snap shot showing the 1,o~r1,1 tajctory plot forthe period that covers 30000 iterations before the ceils re coupled. (k) A Saap shot showing he 2,0~1,1 Sajctry pl Tor te period that overs 30000 iterations bepnning a 600000 eration alr the ely are eoupled. three differential equations only: fea [« tut (a we (12 3) jae-yte = By. aay As can easily be seen, the equations in (12)-(14) define an isolated cell. Here, W specifies the self-feedback connection ‘weight do,» of an isolated cell, hence the condition of W ¢ D yields the double-scroll regime. This proves the necessity.0 1 2 ° 1 2 | (2.2309,2.2941) (2.2499,2.1098) 5 5 (-2.2297 ,-2.2288) ‘xI.d (72-1522 ,-2.2907) xi. (2-2094,2,2640) 2.2307,2.2209) 3 3 Gama aa Ga a Siniar condos on W for oer coherent mods in eB Recognition of Lines wa Phase Schronzation ‘obtained in a same way by considering the J —a.,. bifurcation In the simulations done for the above 3 x 3 network, diagram. it was also observed that i) the complete and partial phaseIEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS: NDAMENTAL THEORY AND APPLICATIONS, VOL. 42, NO. 10, OCTOBER 1995 0 1 2 0 1 2 ® ° <2.2326 ,2.2366) £3.0866,3.0819) 2 4g g g (2.a5a. 2 -Ba7BD xd aera 1 aoIRD xa © © ¢3.1504,3.1311) ¢9.118,3.1117) ° ° § g g 1.4219, -1-4412) xd, ® @ Fig, 12. Fast synchronization depending onthe extemal inputs. The templates are Az — Bra) 3 x 3 image conning a ine ‘wth slope m = +1 represents the external inputs. Black and white corespond tothe 2x0 aad OL levels, respectively. (D) Fist Sap shot showing the rsa taecory plot forthe period that covers 3000) teraons before the calls are coupled. (2) Second Sap shot showing the 11.120 Najectory plot forthe period tat covers 30000 erations afer the cells are coupled. (8) Third Snap shot showing thet; 1—r29 tajecty pt forthe period that covers 30000 iterations beginning at 30000 eran ater the cells are coupled. () Foun snap shot showing the 3 1—r2,0 tajectory plot for the penod that covers 30000 iterations bepnming at 60000R stration after the cells are coupled syneronizations in the coherent double scroll mode can survive The relations between the organization of phase clusters even under time-invariant inputs, ii) phase synchronization and the type of feedback templates, and between phase syn- settiing-time depends on the inputs presented, and iii) in a chronization settling-time and patterns fed via external inputs, partial synchronization regime, the organization of clusters of provide new potentials for pattern recognition. A simple line synchronized cells can be determined by suitable choices of detection application of the above observed phenomena is the feedback templates. presented below.KAVASLAR AND GUZELIS A COMPUTER. ASSISTED INVESTIGATION OF A 2-D ARRAY OF CHUA'S CIRCUITS ms 0 1 2 o 1 2 (2.6628 2.5451) (2.6628 ,2.725) (“LUTE =T. 9652) x00 (1072-1. 711) 0-0 ” © (2.5804 ,2.5706) (2.5804 ,2.584) (HT, 6835 ,-1, 8933) x7 (1, 8839-1. 8802) x00 @ @ Fig. 13. Fast synchronization of the horizontal neighboring cells foe the templates Ax — Be. (a) 3 3 image composed of hoxizonal ines represents the external impute Blick, white, ad the thd gray Tone comespo Yo the rer, 0.10 and 0.0 levels, respectively. (B) Asap shot showing they 1) tajetoy plot forthe period that covers ADCGO iterations afer the eels are Coupled. (6) A snap shot showing the xn ro trajectory pot for the period tat covers 30000 eatonsbegianing at 3000 Iteration afer the cells ae couples. (0) Asap shot showing he,0-T},0uajectry ple To the pied that covers 30000 rations beginning at 60000 eration afer the cells re coupled. fe} Asp sit showing the 0-216 tajectoy lat oe te period that covers 30000 iterations Bepinning at 900th iteration after the cells are couple, In our experiment, the image in Fig. 10(a) is fed via the Different snap shots showing the evolution of an a2 extemal inputs to a 3 x 3 network having the templates trajetory plot associated with wo cells along the diagonal fous 00 : o | (with slope m = —1) are given in Fig. 10(b)-(e). Bach snap Ac=]0 0.74 0 |, Be=|0 005 0]. (15) shot covers a period of time requiring 30000 iterations. The oo os oo Q frst snap shot shows the period before the cells are coupled™ TPSE TRANSACTIONS ON CIRCUTTS AND SYSTEMS: FUNDAMENTAL THEORY AND APPLICATIONS, VOL. 42 NO. 10, OCTOBER 1995 : a.6026,2.5990) 2.2619,2.5094) i i Riese on mars “23a a 6 © . ca.as7s.2.5199) | <2.4629,2.5279) 3 i Goma so Gers meee o o : 2.6004,2.4575) «2.5083 2.2797> g g 17968, - 1, 552B> 0-0 ‘a-8046,-2-0025 3o08 o « Fig 13. (Continued) (A snap shot showing the 70,11, jerry pot for the peti that covers 30000 iterations after the cells are coupled. (g) A sap shat showing tie 20,111, ajectory plot forthe period that covers 30000 iterations beginning st 530000 ate the cols are coupled.) A snap sot showing they": Uajectry ple or the period that covers 30000 eran’ ieginaing at 60000 teraon afer the ces are couple. () A shap shot showing the zo, trajectory pl forthe period tat Covers 3008 iterations beginning a 9000 testi after theese are coupled.) A sap shot showing We 79,0-ro.1tajetory lot forthe ped tat covers 300K eration ater the ells are couple. (k) A soap sbt showing the x0, Uajctry pot for the period that covers 30000 iterations beginning at 00D teraon after the cli are couples, ‘After that instant, 60.000 iterations were sufficient for the cells become synchronized after 60.000 iterations. Any pair of cells €(0,0) and C(1,1) to become synchronized. As shown in that are not lower-right or upper-left neighbors of each other Fig, 10(0-(i), not only cells along the diagonal (with slope have never been observed to synchronize (see Fig. 104j) and ‘mm = =1) but also cells along any line with slope m = 1 (k)). The reason is that such cells are uncoupled for theKAVASLAR AND GOZELIS: & COMPUTER ASSISTED INVESTIGATION OF A 2D ARRAY OF CHUA'S CIRCUETS ns template Ag. The same experiment was repeated for the image in Fig. 11(@). As can be seen from Fig. 11(b)-(e), 90000 iterations (30.000 per snap shot) were not sufficient even for the cells along the diagonal (with slope m = —1) to become synchronized. The results obtained in other experiments done for different input images show that the phase synchronization setting-time becomes minimum for the lines with slope m = =1. Such a fast stimulixdependent synchronization indicates the existence of a line (with slope m = —1) over the pixels corresponding to the synchronized cells. ‘This application has been extended to recognize other patterns in the images. It was observed that the following A; —B; and Ag — Bg templates can be used for recognizing lines with slope m = +1 and horizontal lines, respectively: Op 0s 013 OG O00 o 07 0 |, Be=|0 005 of. 16) Os a0 eo) Dp oe 0 o 0 oO 0 0 0 As= 0.20 0.60 020], Bs=|o 0.10 0]. (17) or 0) oe 0 For the Ay ~ B; templates, cells along the diagonal with slope m = +1 become synchronized quite rapidly if the input image has a diagonal line with slope m = +1 (see Fig. 12). As illustrated by Figs. 13(6)-()), horizontal neighboring cells become synchronized very rapidly (for the As By templates) for the input image of Fig. 13(a) that is composed of three horizontal lines, An observation not included in the figures is that the synchronization is also achieved among cells along any line ofthe input image. In Fig. 13() and ¢k), it can be seen that no synchronization is obtained between cells belonging to different lines, IV. Conc.usion The results obtained by computer simulations on the 2-D array of f(x) ~ # coupled Chua's circuits give insight into the dynamics of the array. Our empirical analysis of mode and phase coherency and bifurcations as related to the connection ‘weights and inputs can be used for further research on high- dimensional chaos and signal processing by chaotic arrays. Tt ccan also provide a basis for theoretical studies on the complex dynamics of 2-D arrays of Chua's circuits. ACKNOWLEDGMENT ‘The authors would like to thank L. , Chua for his valuable comments, REFERENCES (0) C. Guzelig, "Chae cellar neural otworks made of Chus's eeu” 4 Oua’s Cire A Paradigm for Chaos, R. N. Madan, Bd, Word Scientific Series on Nonlinear Science, series By vl. 1. Singapore Wor Scent, 1993, pp. 952-961, (2) V.N-Belykh N.N. Veibev, Lj. Kocarey and LO. Chua, “On cote synetronivtion in liner aay of Chua's ici,” in Chua's Creat A Paradigm for Chane, RN. Madan, Ba, Word Seenie Series om Nonlinear Science, sces B, wo. I. Singapre: Werld Scene, 1993, pp. 325-338 IRM. Dabrowski, W. R. Dabeowshi, and M. J. Oporalk, “Dyoamic ‘henomens in shan interconectons of Chua's cteuts,” IEEE Trans ‘Gets Syst vo. 40, pp. 858-471, Nov, 1983 1MCT‘Oporaie, A.M, Dabrowski and W.R. Dabrowski, “Hyperchas slaserng 0d cooperative phenomena in CNN arays composed of ‘Shao eeu tn Poe, Third IEEE Int. Workshop on Cellar Neural [Nemorks and Their Applicat, CNNA "98, Rome, 1994, pp. 315-320, (CGazty and LO. Chua, “Subilty analysis of geneaied cellular cual nctwork" J Cre Theory Applicat. pp. 1-3, 193. [6 K. Kaneko, "Csering, coding, sch. hcarchialetdring and contrat in a network of chao clemeats" Physica D, voll, pp. 137-172. 1990, WW. iceman, “Tusial on neurobiology frm single nerons to brain chaos." In. J Bf, and Chas, vol. 3-20. 3, pp 451-482, 1992 Baird, MW. Hisch and F. Bockman, "A neal network associative memory for handwiten character recognition using mulipe Chua tractors” IEEE Trane. Cire Syat v0.40, pp. 667-674, 193, {9} LO. Chis, “The genesis of Chua's ciculc” Arc Fle, Uberraguns, ok 46, pp. 250-257. 1992, {10} K" Muraland M. Lakshmanan “Esfets of sinasoidsl excitation om the CCh'scireit IEEE Trans, Circuit Sat vl. 39, pp. 267-270, 19 [11] 1M Chur and LO, Chan, “& CMOS 1C nonlinear resistor for Chua's ‘uit FEEE Trans. Cine Syt vol 39, pp. 985-998, 1992 [12] MP. Kennedy. “Robust op amp implementation of Chua's circuit.” Freguens, voli, pp 6-80, 19 113) 0" Chon and’ Le Yang, “Cellar neal aerwoeks: Theory and pplication” IEEE Trane. Crt Spt. vol 38. pp. 1287-1272, 1988. 114) WPecezhunuzu, V- Perer Vile, and L. O. Chon “Traveling wave front and ie fale in a one dimensional aray of Chua's cies in Cho's Circuit: A Paradigm for Chaos, RN. Madan, Ea, World Scientific Serien-on Noninear Science, sexes B, vol. Sigupore ‘World Scientii, 19, pp. 336-380. Autowaves for Image processing ona two dimensional CNN a. ‘ay of exehable nonlinear Chua's cults: Flat and wrinkled abn, IEEE Trans, Circus Sut, vo. 40, pp 174-181, Mar. 193 8 io ‘5 nm ast atin Kavaslar recived the BS. and MS. apres ‘in leercal engineering from Isanbal Techical University, Turkey i 1991 and 1995, respectively. Hes correctly 2 PhD, student in sistnal en fsering at Istanbul Tesi! University Tn 1992, he joined Siemens Company, intanbl 4s A menier of the Power Hecuenics Division He won the 1989 Turkish championship ip tn sailing. His research interests ares di signal processing, neural networks and compster vison Caney Gael tscved the BS. M.S. and PD. Seqresinclectal engineering rom nb Tec tical Univers Turkey in T9618, and 1988, perc The worked ot tanbs! Technical University a “Teaching Asta between ORD an 1980, Pos ‘Al 1589 fo Apo TOD, be wah isting Re ‘altho and Lecter with be Depa of ls teal Enpnering snd Computer Slences, Univer Fe Siy’ot Gitorma a Bere. He jied Iau {etl Univer in 1991 where Be nan a Assvite Professor inthe Faculy of Eleccal Electronics Engineering, His esearch ineess ae in nslinear cits and systems, neural networks, nd Seta processing