1: %format=latex

2: 

3: \documentstyle[12pt]{article}

4: \renewcommand{\baselinestretch}{1.5}

5: \oddsidemargin=0cm

6: \textwidth=16cm

7: \textheight=23cm

8: \topmargin=0cm

9: \headheight=0cm

10: \headsep=0cm

11: \topskip=0cm

12: 

13: \newcommand{\ee }{^{\mbox{\scriptsize EE}}}

14: \newcommand{\ei }{^{\mbox{\scriptsize EI}}}

15: \newcommand{\ie }{^{\mbox{\scriptsize IE}}}

16: \newcommand{\ii }{^{\mbox{\scriptsize II}}}

17: 

18: \newcommand{\E }{^{\mbox{\scriptsize E}}}

19: \newcommand{\I }{^{\mbox{\scriptsize I}}}

20: 

21: \newcommand{\low }{_{\mbox{\scriptsize low}}}

22: \newcommand{\sn }{_{\mbox{\scriptsize sn}}}

23: \newcommand{\hopf }{_{\mbox{\scriptsize hopf}}}

24: 

25: \newcommand{\eqdef}{\stackrel{\rm def}{=}}

26: 

27: \begin{document}

28: 

29: \input{psfig}

30: 

31: \begin{titlepage}

32: \centerline{\huge COVARIANCE PLASTICITY}

33: \vglue .5cm

34: \centerline{\huge AND REGULATED CRITICALITY}

35: \footnotetext

36: {Supported by the Jean and H\'{e}l\`{e}ne Alfassa fund

37: for research in Artificial Intelligence,

38: Office of Naval Research contract N00014-91-J-1021,

39: National Science Foundation contract DMS-9217655,

40: and ARL contract MDA972-93-1-0012.}

41: \vglue 1cm

42: \centerline{\bf Elie Bienenstock}

43: \vglue .2cm

44: \centerline{Division of Applied Mathematics}

45: \centerline{Brown University}

46: \centerline{Providence RI 02912}

47: \centerline{USA}

48: \centerline{and CNRS, Paris, FRANCE}

49: \vglue .2cm

50: \centerline{elie@dam.brown.edu}

51: \vglue .5cm

52: \centerline{\bf Daniel Lehmann}

53: \vglue .2cm

54: \centerline{Department of Computer Science}

55: \centerline{Hebrew University}

56: \centerline{Jerusalem}

57: \centerline{ISRAEL}

58: \vglue .2cm

59: \centerline{lehmann@cs.huji.ac.il}

60: \vglue .5cm

61: \centerline{January 1995}

62: 

63: \end{titlepage}

64: 

65: \vglue 1cm

66: \begin{abstract}

67: We propose that a regulation mechanism

68: based on Hebbian covariance plasticity

69: may cause the brain to operate near criticality.

70: We analyze the effect of such a regulation

71: on the dynamics of a network with excitatory and inhibitory neurons

72: and uniform connectivity within and across the two populations.

73: We show that, under broad conditions,

74: the system converges to a critical state

75: lying at the common boundary of three regions in parameter space;

76: these correspond to three modes of behavior:

77: high activity, low activity, oscillation.

78: \end{abstract}

79: 

80: \newpage

81: 

82: \section{Introduction}

83: \label{intro}

84: 

85: That evolved brains are highly sensitive organs is an everyday observation.

86: Viewed as a dynamical system,

87: a brain may be said to be {\em unusually} susceptible to perturbations

88: and initial conditions.

89: This leads one to ask whether

90: brains may be operating near some form of instability, or criticality,

91: a hypothesis related to the notions

92: of computation at the edge of chaos (Langton 1990)

93: and self-organized criticality (Bak et al. 1987).

94: In this paper we propose that

95: while most regulation mechanisms at work in the brain

96: act according to a classical homeostasis schema,

97: i.e., have a stabilizing effect,

98: an opposite effect could result from the regulation of synaptic weights

99: by a specific form of Hebbian covariance plasticity.

100: Such a regulation may bring the system near criticality.

101: We suggest that regulated criticality may be the mechanism

102: whereby sensitivity is {\em maintained} throughout life

103: in the face of ongoing changes in brain connectivity.

104: 

105: Hebbian synaptic plasticity (Hebb 1949) plays an important role

106: in the development of the nervous system,

107: and is also believed to underlie many instances of learning in the adult.

108: A {\em covariance rule} of Hebbian plasticity roughly states

109: that the change in the efficacy of a given synapse

110: varies in proportion to the covariance

111: between the presynaptic and postsynaptic activities.

112: As noted by many authors

113: (e.g. Sejnowski 1977a, 1977b; Bienenstock et al. 1982;

114: Linsker 1986; Sejnowski et al. 1988),

115: a covariance-type rule is preferable to a rule that uses the mere product

116: of pre- and post-synaptic activities

117: because the covariance rule predicts not only weight increases

118: but also activity-related weight decreases,

119: and as a consequence allows convergence to non-trivial connectivity states.

120: Some forms of covariance plasticity have been shown to be optimal

121: for information storage

122: (Willshaw and Dayan 1990; Dayan and Willshaw 1991; Dayan and Sejnowski 1993).

123: Also, evidence for Hebbian plasticity of the covariance type

124: has been reported in several preparations

125: (Fr\'egnac et al. 1988, 1992; Stanton and Sejnowski 1989;

126: Artola et al. 1990; Dudek and Bear 1992).

127: 

128: We shall investigate,

129: in a simple network including excitatory and inhibitory neurons,

130: the effect of covariance plasticity

131: acting as a mechanism of {\em regulation,} rather than supervised learning.

132: Synaptic modification results

133: in changes---quantitative or qualitative---in the activity

134: that reverberates in the network;

135: these changes in turn cause further modification of the weights,

136: thereby creating a feedback loop between activity and connectivity.

137: Studying this loop as such,

138: i.e., independently from any input and output,

139: we demonstrate that, under fairly general conditions,

140: it causes the network to converge to a critical surface in parameter space,

141: the locus of an abrupt transition between different activity modes.

142: In Metzger and Lehmann (1990, 1994) a similar Hebbian rule

143: has been studied in the context of supervised learning of temporal sequences.

144: 

145: Schematically, the convergence to a critical state can be explained as follows.

146: One mode of behavior of a network

147: including excitatory and inhibitory neurons is oscillation;

148: such behavior takes place if the synaptic weights

149: linking excitatory neurons to each other---we

150: will refer to these as E-to-E weights---are

151: high enough but not too high.

152: Oscillation entails high covariance values,

153: hence, according to the covariance rule,

154: results in further increase of the E-to-E weights.

155: If however these weights are higher than a certain critical value---which

156: depends on other parameters of the system---oscillatory

157: behavior is impossible,

158: hence covariance is low or zero,

159: hence, in accord with the covariance rule used,

160: the E-to-E weights {\em decrease.}

161: As a result, the E-to-E weights stabilize around

162: the critical surface that separates the region of oscillation

163: from the region(s) of steady firing.

164: 

165: Our study is conducted in the simplest type of network

166: that will support oscillatory activity:

167: all synaptic weights of a given type are given identical values,

168: and so are all firing thresholds of a given type.

169: This results in a system with just six parameters---four

170: synaptic weights and two thresholds---and a limited range of behaviors.

171: Essentially, all neurons fire uniformly, either at a constant rate

172: (the number of possible rates of firing is one or two, depending on parameters)

173: or periodically in time.

174: In the {\em thermodynamic,} i.e., large-size, limit,

175: the dynamics of the network is adequately described

176: by a system of differential equations

177: obtained through a classical mean-field approximation.

178: 

179: We first perform a simple bifurcation analysis of this differential system

180: (Guckenheimer and Holmes 1983).

181: We then show that the effect of covariance regulation

182: is to stabilize the parameter state

183: at a surface of transition,

184: where the dynamics exhibits an instability.

185: Such a critical parameter state for a dynamical system

186: may be characterized as {\em degenerate.}

187: A generic, i.e., non-exceptional, state is one

188: where one expects to find the system

189: in the absence of further assumptions.

190: Mathematically, a generic state of a dynamical system is in the {\em interior}

191: of a parameter region corresponding to a given behavior,

192: and the system in such a parameter state

193: is said to be {\em structurally stable;}

194: the set of non-generic parameter states has measure zero.

195: We shall show that a state of higher degeneracy,

196: characterized as a point of intersection of {\em several} critical surfaces,

197: can be achieved by the simultaneous regulation of {\em several} parameters.

198: In the vicinity of that highly degenerate state,

199: the system displays a range of behaviors, including chaos.

200: 

201: The plan of the paper is as follows.

202: In the next section we study the dynamical properties

203: of our simple network---in the differential-equation formulation---with

204: {\em fixed} parameters (synaptic weights and firing thresholds).

205: We characterize the bifurcations

206: which take place at the boundaries between domains

207: corresponding to different modes of behavior.

208: This study is conducted for a {\em reduced} system,

209: where the thresholds are eliminated in such a way

210: as to render the dynamics symmetric about the origin.

211: Section \ref{regulation} describes the regulation equations.

212: Section \ref{reduced} describes the behavior of the regulated reduced system,

213: and Section \ref{full} that of the regulated full system.

214: 

215: \section{The fixed-parameter model}

216: \label{model}

217: 

218: This section describes the dynamics of the model with fixed parameters.

219: We first briefly describe a network consisting of a large number ($2N$)

220: of binary-valued neurons operating under a stochastic dynamics.

221: However, rather than using this network for our study of plasticity,

222: we make a number of simplifications and approximations,

223: leading to a deterministic two-variable differential system

224: with just six parameters.

225: The two variables are the excitatory and inhibitory

226: population averages of cell activity

227: in the $2N$-dimensional model;

228: the six parameters include the four average weights of the synapses

229: within and between these two populations,

230: as well as the average firing thresholds for the two populations.

231: We then study the asymptotic behavior

232: of this differential system for various parameter values.

233: Different types of asymptotic behavior,

234: in different regions of the parameter space,

235: correspond to different {\em phases} of the stochastic system,

236: and we pay particular attention to the {\em bifurcations} of the solutions,

237: where the bifurcation parameters are the synaptic weights---see

238: Schuster and Wagner (1990) and Borisyuk and Kirillov (1992)

239: for a related bifurcation analysis.

240: Bifurcations correspond to {\em phase transitions}

241: in the statistical-physics formulation

242: (the original $2N$-dimensional model).

243: 

244: We consider a fully-connected network

245: of $N$ excitatory and $N$ inhibitory linear-sigmoidal

246: $\{0,1\}$-valued neurons,\footnote

247: {It is not essential

248: that the numbers of excitatory and inhibitory neurons be the same.}

249: operating under a stochastic dynamics.

250: We denote the activity of the $i$-th excitatory,

251: resp. inhibitory, neuron by

252: $x_i\E (t)$, resp. $x_i\I (t)$,

253: with $x_i\E (t), x_i\I (t)\in\{0,1\}$, $i=1\ldots N$,

254: and we denote the synaptic weights by

255: $w_{ij}\ee , w_{ij}\ei , w_{ij}\ie , w_{ij}\ii $, $i,j=1\ldots N$,

256: where $i$ is postsynaptic and $j$ presynaptic,

257: and the superscripts indicate,

258: for each of the two neurons, whether it is excitatory or inhibitory.

259: Thus, for all $i$ and $j$, $w_{ij}\ee $ and $w_{ij}\ie $ are positive or zero,

260: whereas $w_{ij}\ei $ and $w_{ij}\ii $ are negative or zero.

261: 

262: The {\em local field} on excitatory neuron $i$,

263: i.e., the difference between its membrane potential

264: and its firing threshold $h_i\E $,

265: is $g_i\E (t)=\sum_j w_{ij}\ee  x_j\E (t) +

266: \sum_j w_{ij}\ei  x_j\I (t) - h_i\E $.  

267: Similarly, the local field on inhibitory neuron $j$ is

268: $g_i\I (t)=\sum_j w_{ij}\ie  x_j\E (t) +

269: \sum_j w_{ij}\ii  x_j\I (t) - h_i\I $,

270: where $h_i\I $ is the threshold of inhibitory neuron $i$.

271: The network dynamics is defined by:

272: (i) selecting at random, with uniform probability,

273: one of the $2N$ neurons;

274: (ii) computing its local field $g(t)$,

275: of the form $g_i\E (t)$ or $g_i\I (t)$; and

276: (iii) defining the state of the network at time $t+\delta t$

277: to be equal to the state at time $t$

278: except, possibly, for the selected neuron, whose state becomes---or stays---1

279: with probability $\frac{1}{2}(1+\tanh(\beta g(t)))$.

280: Parameter $\beta$ is a fixed non-negative number,

281: an {\em inverse temperature}.

282: The temperature $T=1/\beta$ measures the amount of noise in the system:

283: the higher the temperature, the noisier the dynamics.

284: The update interval is $\delta t = 1/(2N)$,

285: so that each neuron is updated on average once every time unit.

286: This {\em asynchronous} dynamics, of the Glauber type (Glauber 1963),

287: is widely used in statistical-mechanics models;

288: it lends itself to a convenient mean-field approximation (see below).

289: 

290: A system such as the one just described

291: will exhibit a highly diverse range of behaviors,

292: depending on the values of the synaptic weights and firing thresholds.

293: But we now make the much simplifying assumption that synaptic weights

294: and firing thresholds are {\em uniform} across each class.

295: Specifically, for all $i,j=1,\ldots,N$,

296: we assume that $h_i\E =h\E $, $h_i\I =h\I $,

297: $w_{ij}\ee =w\ee /N$, $w_{ij}\ei =-w\ei /N$,

298: $w_{ij}\ie =w\ie /N$, and $w_{ij}\ii =-w\ii /N$,

299: where $h\E $, $h\I $, $w\ee $, $w\ei $, $w\ie $ and $w\ii $ are fixed parameters,

300: and $w\ee $, $w\ei $, $w\ie $ and $w\ii $ are non-negative.

301: The dynamics is thus parameterized by six constants,

302: four synaptic weights and two thresholds;

303: $\beta$ is a mere multiplicative factor common to all six parameters,

304: yet it is convenient to use it as a seventh parameter.

305: Unless otherwise mentioned, $\beta$ will be 1.

306: 

307: Due to this uniformity assumption,

308: all neurons in any of the two populations experience the same field

309: at any given time.

310: This system exhibits a limited number of fairly simple behaviors,

311: of which Figure 1 is an example.

312: This figure shows the time variation of

313: $\langle x_i\E (t)\rangle$ and $\langle x_i\I (t)\rangle$,

314: the {\em average} activation levels across

315: the excitatory and inhibitory populations.

316: In this example, parameters are:

317: $N=70$, $w\ee =12$, $w\ie =8$, $w\ei =10$, $w\ii =2$,

318: $h\E = 1$, $h\I = 3$.

319: One unit on the time axis corresponds to $2N$ updates,

320: so that each neuron is updated, on average, once every time unit.

321: For these parameter values, the system {\em oscillates.}

322: Note that the oscillation is not perfectly regular, a finite-size effect.

323: Note also that the inhibitory activity lags somewhat behind the excitatory activity:

324: the excitatory neurons first trigger the inhibitory ones,

325: which in turn extinguish, for a while, the excitatory population.

326: \vglue .4cm

327: \centerline{\em (Insert Figure 1 around here)}

328: \vglue .4cm

329: The presence of oscillations and the amplitude and shape of the waveform

330: depend on the various parameters.

331: However, rather than pursuing this study of the stochastic system,

332: we shall consider the approximation that obtains

333: in the {\em thermodynamic limit},

334: that is, when $N\rightarrow\infty$.

335: The update interval $\delta t=1/(2N)$ then goes to 0

336: and so does each individual synaptic weight.

337: Straightforward approximations (Rubin 1988; Schuster and Wagner 1990)

338: then lead to a continuous-time differential system

339: for the population averages of the excitatory and inhibitory activation levels,

340: which we denote, respectively, by $s$ and $\sigma$:

341: \begin{equation} \left\{ \begin{array}{l}

342: \dot{s}(t) = .5 - s(t) + .5\tanh[\beta(w\ee s(t)-w\ei \sigma(t)-h\E )]\\

343: \dot{\sigma}(t)=.5- \sigma(t) +.5\tanh[\beta(w\ie s(t)-w\ii \sigma(t)-h\I )].

344: \end{array} \right. \label{sys_full} \end{equation}

345: Note that the variables $s(t)$ and $\sigma(t)$

346: remain at all $t$ within the interval [0,1].

347: When $\beta=0$

348: system \ref{sys_full} has a unique attractor, $(s,\sigma)=(.5,.5)$.

349: Indeed, in the high-temperature limit,

350: all neurons act independently of each other

351: and fire with probability .5 at each time.

352: 

353: We shall now make a last simplification,

354: whose purpose it is to render $(.5,.5)$ a fixed point---though

355: not necessarily stable---at {\em all} temperatures

356: and for all values of the synaptic weights.

357: This is easily achieved by letting the thresholds $h\E $ and $h\I $

358: be determined by the synaptic weights as follows:

359: \begin{equation} \begin{array}{l}

360: h\E  = .5 (w\ee -w\ei )\\

361: h\I  = .5 (w\ie -w\ii ).

362: \end{array} \label{chvar} \end{equation}

363: It is then convenient to adopt the change of variables:

364: $s\mapsto s-.5$, $\sigma\mapsto \sigma-.5$,

365: and system \ref{sys_full} becomes:

366: \begin{equation} \left\{ \begin{array}{l}

367: \dot{s}(t) = -s(t) + .5\tanh[\beta(w\ee s(t)-w\ei \sigma(t))]\\

368: \dot{\sigma}(t) = -\sigma(t) + .5\tanh[\beta(w\ie s(t)-w\ii \sigma(t))].

369: \end{array} \right. \label{sys_red} \end{equation}

370: In \ref{sys_red}, the variables $s$ and $\sigma$

371: are in the interval $[-.5,+.5]$,

372: and the only parameters left are the four synaptic weights

373: and the inverse temperature.

374: For all parameter values, the origin is a fixed point of system \ref{sys_red}.

375: A different position for the fixed point

376: could be obtained with an appropriate modification of equations \ref{chvar},

377: yet in the current version the fixed point is also a center of symmetry.

378: For the moment,

379: this hard-wired symmetry should be regarded as an {\em ad-hoc} device,

380: whose purpose is to make the analysis more convenient.

381: We shall refer to system \ref{sys_full} as the {\em full} system,

382: and to system \ref{sys_red} as the {\em reduced} system.

383: We shall see in Section \ref{full}

384: that under appropriate regulation

385: the two systems behave similarly.

386: 

387: We now discuss some important properties of the reduced system,

388: system \ref{sys_red} (see also Rubin 1988).

389: Consider first Figure 2a (phase diagram), which shows

390: four trajectories of the state $(s(t),\sigma(t))$;

391: the starting points of these trajectories are indicated by triangles.

392: The parameters (synaptic weights) used in this example are identical

393: to those used in Figure 1, i.e.,

394: $w\ee =12$, $w\ie =8$, $w\ei =10$, $w\ii =2$.

395: As expected, the asymptotic behavior is {\em periodic;}

396: there is a limit cycle which attracts all points of the square $[-.5,.5]^2$,

397: except the unstable equilibrium $(0,0)$.

398: Motion is counterclockwise,

399: for, as mentioned above, $\sigma(t)$ lags behind $s(t)$.

400: 

401: In addition to these four orbits, Figure 2a shows two curves,

402: the $s$- and $\sigma$-{\em nullclines} for system \ref{sys_red}.

403: These are the loci of the points $(s,\sigma)$ such that $ds/dt$,

404: resp. $d\sigma/dt$, vanish.

405: The equations for the $s$- and $\sigma$-nullclines

406: are easily seen to be, respectively:

407: \begin{equation}

408: \sigma = \frac{1}{w\ei }(w\ee s - T\tanh^{-1}(2s)),

409: \label{snull}\end{equation}

410: \begin{equation}

411: s = \frac{1}{w\ie }(w\ii \sigma + T\tanh^{-1}(2\sigma)).

412: \label{signull}\end{equation}

413: The $\sigma$-nullcline is an increasing sigmoid-shaped curve,

414: whereas the $s$-nullcline generally has the shape of an `S' lying on its side.

415: Of particular interest are the intersection points of the two nullclines;

416: these are the {\em fixed points} of the dynamics.

417: In the case illustrated in Figure 2a, the only intersection is $(0,0)$,

418: an unstable equilibrium.

419: Trajectories intersect the $s$-, resp. $\sigma$-, nullcline

420: in a direction parallel to the $\sigma$-, resp. $s$-, axis.

421: 

422: The study of the nullclines is of interest

423: because it is often possible

424: to predict how a parameter change will affect the dynamics of the system

425: by reasoning about how the nullcline diagram will change;

426: the bifurcation we shall be mostly interested in

427: is associated with a conspicuous change in this diagram.

428: Note that the $s$-nullcline is affected by parameters $w\ee $ and $w\ei $,

429: whereas the $\sigma$-nullcline is affected by parameters $w\ii $ and $w\ie $.

430: \vglue .4cm

431: \centerline{\em (Insert Figure 2 around here)}

432: \vglue .4cm

433: Let us consider first the changes brought about

434: by letting parameter $w\ee $ grow,

435: starting from the point $w\ee  = 12$

436: for which the system oscillates;

437: other parameters are unchanged.

438: When $w\ee $ grows, the slope of the central,

439: quasi-linear, part of the $s$-nullcline increases

440: (see equation \ref{snull});

441: that part of the curve rotates about the symmetry center (0,0).

442: As a result, the peak of the $s$-nullcline to the right approaches

443: the upper part of the sigmoid-shaped $\sigma$-nullcline,

444: and the minimum of the $s$-nullcline to the left approaches

445: the lower part of the $\sigma$-nullcline.

446: Eventually, at a certain critical value $\hat w\ee \sn $

447: (subscript `sn' stands for `saddlenode'---see below),

448: the two curves become tangent to each other.

449: This happens in two points at once,

450: near the upper right-hand corner and near the lower left-hand corner,

451: due to the symmetry of the system.

452: This situation is depicted in Figure 2b:

453: $w\ee $ is exactly equal to the critical value $\hat w\ee \sn $

454: (with parameters as above, $\hat w\ee \sn \approx 14.22$),

455: and the nullclines are just tangent to each other.

456: 

457: When $w\ee $ grows a little further,

458: each point of contact splits into two intersection points,

459: of which one is an attractor.

460: Figure 2c shows this situation, with $w\ee =15$,

461: somewhat above the critical value $\hat w\ee \sn $.

462: Four trajectories are shown, in addition to the two nullclines.

463: The system has five fixed points,

464: three unstable ones and two stable ones (attractors).

465: Only the stable fixed points are of interest to us;

466: they are very near the upper right-hand

467: and lower left-hand corners of the square,

468: corresponding to high,

469: respectively low, excitatory and inhibitory activities.

470: 

471: The bifurcation occurring at $\hat w\ee \sn $

472: is of the {\em saddlenode} type.

473: It results in a drastic change of behavior of the system:

474: the periodic attractor disappears

475: and is `siphoned' into the two new point attractors.

476: These two points attract the entire square,

477: except a set of measure 0

478: which includes the three unstable fixed points.

479: Thus, altough this bifurcation

480: is caused by a mere {\em local} change,

481: namely the intersection of the nullclines,

482: it results in a reorganization of the dynamics

483: that is both abrupt and {\em global.}\footnote

484: {As mentioned, {\em two} distinct saddlenode bifurcations

485: take place simultaneously.

486: Such a double bifurcation is not generic;

487: it occurs here due to the symmetry that we introduced

488: when reducing system \ref{sys_full} into system \ref{sys_red}.}

489: 

490: Having described the breakdown of oscillations

491: when parameter $w\ee $ is increased,

492: we now consider the opposite change, that is,

493: we let $w\ee $ decrease.

494: This results in a decrease of the slope of the central,

495: increasing, portion of the $s$-nullcline (equation \ref{snull}).

496: Eventually, the curve becomes monotonically decreasing;

497: this does not alter the number of intersections of the nullclines,

498: point $(0,0)$ remaining the sole equilibrium.

499: However, the amplitude of the limit cycle decreases along with $w\ee $.

500: The cycle eventually collapses to a point;

501: the equilibrium $(0,0)$ has then become stable.

502: This can be seen in a linear stability analysis

503: of system \ref{sys_red} around point $(0,0)$.

504: It is easily shown that,

505: in case there are two complex conjugate eigenvalues,\footnote

506: {The condition for this is $4w\ei w\ie >(w\ee +w\ii )^2$.}

507: the real part of these eigenvalues is negative if and only if $w\ee <w\ii +4T$.

508: Thus, $w\ii +4T$ is a critical value for parameter $w\ee $.

509: We define $\hat w\ee \hopf \eqdef w\ii +4T$

510: (with the current parameter setting, $\hat w\ee \hopf =6$).

511: The change of behavior occurring at $\hat w\ee \hopf $

512: is a {\em normal}\footnote

513: {That is, supercritical.

514: However, for very large values of $w\ie $,

515: the bifurcation is subcritical---see footnote 5.}

516: Hopf bifurcation.

517: 

518: So far, we studied the behavior of system \ref{sys_red}

519: for different values of parameter $w\ee $, all other parameters being fixed.

520: In other words, we described the system's behavior

521: on a particular 1-dimensional subspace

522: of the 4-dimensional parameter space.

523: We now extend this study to a 2-dimensional subspace, the $(w\ee ,w\ie )$ plane.

524: Figure 3a is the bifurcation diagram of system \ref{sys_red} in that plane,

525: with other parameters as before ($w\ei =10$, $w\ii =2$).

526: This diagram shows three distinct regions,

527: corresponding to three different attractor configurations;

528: unstable fixed points and unstable limit cycles are ignored in this diagram.

529: In the middle region---which we call region $\cal P$,

530: for {\em Periodic}---the system oscillates.

531: The boundary of this region to the right

532: is the saddlenode bifurcation curve, which we denote $\cal S$;

533: as discussed above,

534: the rightmost region has {\em two} point attractors,

535: and we call it region $\cal T$.

536: The leftmost region, which we call $\cal O$,

537: has only one point attractor, the center of symmetry $(0,0)$;

538: it is separated from region $\cal P$ by the Hopf bifurcation curve,

539: a vertical line of equation $w\ee =\hat w\ee \hopf $.

540: The curve in the lower left of the diagram,

541: separating region $\cal O$ from region $\cal T$,

542: is the locus of a {\em pitchfork} bifurcation.

543: This bifurcation diagram, obtained for one particular set of values

544: of the parameters $w\ei , w\ii $ and $\beta$,

545: is representative of the general case.\footnote

546: {It is however simplified in two ways.

547: First, the transition from region $\cal P$ to region $\cal T$

548: is of the saddlenode type only for large enough values of $w\ie $;

549: this range of values corresponds roughly

550: to the straight portion of curve $\cal S$ (Figure 3a).

551: To see why this is so, consider again Figure 2b,

552: the nullcline diagram at the bifurcation, with $w\ie = 8$.

553: Note that the points of contact between the nullclines

554: appear near the corners of the square, far from the origin;

555: this is due to the fact that $w\ie $ is relatively large,

556: hence the slope of the $\sigma$-nullcline at the origin

557: is larger than the slope of the $s$-nullcline.

558: The bifurcation is then of the saddlenode type, as described.

559: If however $w\ie $ is small,

560: hence so is the slope of the $\sigma$-nullcline at the origin, 

561: the transition from $\cal P$ to $\cal T$ as $w\ee $ is increased

562: takes place differently.

563: A pair of intersection points between the nullclines

564: first split off {\em from the origin;}

565: these are unstable equilibria.

566: As $w\ee $ increases,

567: these two equilibria move away from the origin,

568: while remaining inside the large stable limit cycle.

569: At a certain critical value for $w\ee $ they become stable---a

570: (double) subcritical Hopf bifurcation---and

571: almost immediately thereafter the large limit cycle disappears.

572: Thus, the transition from region $\cal P$ to region $\cal T$

573: really takes place in two steps,

574: giving rise to a {\em three-attractor} behavior:

575: the system has one large limit-cycle attractor

576: {\em as well as} two point attractors,

577: the latter being inside the cycle.

578: The region of the $(w\ee ,w\ie )$ plane where this behavior takes place

579: is a strip extending along the lower, curved,

580: part of the $\cal P/\cal T$ boundary;

581: it is too narrow to be seen in Figure 3a.

582: (With parameters $w\ei $ and $w\ii $ as above and $w\ie =2.75$,

583: the three-attractor behavior occurs for $w\ee $ between $8.993$ and $9.030$.

584: For some other values of $w\ei $ and $w\ii $ this behavior does not occur at all,

585: and the transition from $\cal P$ to $\cal T$ is always of the saddlenode type.)

586: For the purpose of this paper (see footnotes 8 and 11)

587: it is important to note that the point attractors

588: appear either {\em exactly} or {\em almost} at the same time

589: as the periodic attractor disappears.

590: The second approximation in the bifurcation diagram,

591: mentioned only for the sake of completeness,

592: concerns the $\cal O$-to-$\cal P$ transition.

593: This is generally a smooth, supercritical, Hopf bifurcation.

594: However, as mentioned in footnote 4, this Hopf bifurcation becomes subcritical

595: for very large values of $w\ie $.

596: There is thus a narrow region to the left of the bifurcation line $w\ee =\hat w\ee \hopf $

597: where the limit-cycle attractor coexists with the point attractor (0,0);

598: for instance, at $w\ie =100$, the width of this region is $\approx 0.63$.}

599: \vglue .4cm

600: \centerline{\em (Insert Figure 3 around here)}

601: \vglue .4cm

602: In sum, the $(w\ee ,w\ie )$ bifurcation diagram for system \ref{sys_red}

603: is characterized by a central periodic-attractor region,

604: a large vertical patch extending to $+\infty$ in the $w\ie $ direction

605: (phase $\cal P$),

606: flanked by point-attractor regions on each side (phases $\cal O$ and $\cal T$).

607: The transition from $\cal P$ to $\cal T$ is abrupt ($\cal S$ line),

608: while the transition from $\cal O$ to $\cal P$ is smooth.

609: As mentioned in the Introduction, system \ref{sys_full}---the full system---is

610: not amenable to such a thorough analysis;

611: however, we shall see in Section \ref{full}

612: that the two systems behave in much the same way

613: under the plasticity rules that we shall now introduce.

614: 

615: \section{The regulation equations}

616: \label{regulation}

617: 

618: Whereas in the previous section the synaptic weights $w\ee $ and $w\ie $

619: were fixed parameters, they will now be made to evolve.

620: Their evolution will obey a Hebbian covariance rule, hence be a function

621: of second-order temporal averages of the dynamic variables $s$ and $\sigma$.

622: Synaptic plasticity creates a {\em regulation loop:}

623: changing the parameters affects the dynamics of the system,

624: which in turn alters the second-order moments of $s$ and $\sigma$.

625: Formally, the regulation is implemented by

626: introducing additional differential equations, coupled to system \ref{sys_red}

627: (or to system \ref{sys_full}---see Section \ref{full}).

628: The rate of change of $w\ee $ and $w\ie $ will typically be

629: several orders of magnitude slower than that of $s$ and $\sigma$.

630: 

631: Let us first define, for any function of time $r(t)$,

632: a moving time average:

633: \begin{equation}

634: \bar{r}(t) = \rho \int_{-\infty}^t r(u) e^{\rho (u-t)} du.

635: \nonumber\end{equation}

636: Parameter $\rho$ is a positive constant, physically an inverse time;

637: the larger $\rho$, the narrower the averaging kernel.

638: Equivalently, $\bar{r}(t)$ may be defined by a differential equation,

639: more convenient for simulation purposes:

640: \begin{equation}

641: \frac{d\bar{r}(t)}{dt} = \rho (r(t) - \bar{r}(t)).

642: \nonumber\end{equation}

643: Consider now, with reference to the original stochastic model (Section \ref{model}),

644: the {\em instantaneous covariance} between two excitatory neurons $i$ and $j$,

645: defined as:

646: $c_{ij}\ee (t) \eqdef (x_i\E (t)- \bar x_i\E (t))(x_j\E (t) - \bar x_j\E (t))$.

647: If we take the {\em population average} $\langle c_{ij}\ee (t)\rangle$

648: of this instantaneous covariance,

649: we obtain, in the thermodynamic limit $N\rightarrow\infty$,

650: the instantaneous variance of $s(t)$:

651: \begin{equation}

652: c\ee (t) \eqdef (s(t) -\bar{s}(t))^2.

653: \label{covee}\end{equation}

654: 

655: It is this quantity $c\ee $ that we use to regulate

656: the excitatory-to-excitatory synaptic weight $w\ee $.

657: The regulation equation is linear in $c\ee $:

658: \begin{equation}

659: \frac{dw\ee (t)}{dt} = \varepsilon\ee  ( c\ee (t) - \theta\ee  ).

660: \label{regee}\end{equation}

661: Parameters $\varepsilon\ee $ and $\theta\ee $ are positive.

662: Note that the quantity $c\ee (t)$ is always non-negative;

663: the term $-\theta\ee$ is therefore necessary

664: to allow for decreases of $w\ee $.

665: 

666: We shall also consider a regulation for $w\ie $,

667: the synaptic weight from excitatory to inhibitory neurons,

668: although this regulation will play a less important role than that of $w\ee $. 

669: The modification rule for $w\ie $ has the same form as equation \ref{regee},

670: yet it uses the excitatory-to-inhibitory instantaneous covariance,

671: defined as:

672: \begin{equation}

673: c\ie (t) \eqdef (s(t) -\bar{s}(t)) (\sigma(t) -\bar{\sigma}(t)).

674: \label{covie}\end{equation}

675: The regulation equation for $w\ie $ then reads:

676: \begin{equation}

677: \frac{dw\ie (t)}{dt} = \varepsilon\ie  ( c\ie (t) - \theta\ie  ).

678: \label{regie}\end{equation}

679: In equation \ref{regie}, $\theta\ie $ is a positive constant,

680: as $\theta\ee $ in equation \ref{regee}.

681: However, the modification rate constant $\varepsilon\ie $ is negative.

682: The main reason for this will be given in the next section;

683: for now, note that this choice is consistent with the spirit of Hebb's principle,

684: for, when considered {\em postsynaptically} to the target neuron,

685: the effect of synapse reinforcement if that target neuron is inhibitory

686: is the opposite of the effect obtained if the target neuron is excitatory.

687: 

688: \section{Behavior of the regulated reduced system}

689: \label{reduced}

690: 

691: This section describes the behavior of the regulated reduced system.

692: We demonstrate that each of the two regulation loops

693: introduced in Section \ref{regulation},

694: when acting separately,

695: brings the system to the critical surface $\cal S$,

696: the locus of an abrupt phase transition (saddlenode bifurcation).

697: We then examine the behavior of the system

698: with the two regulation loops active simultaneously;

699: we show that under some conditions

700: the state converges to a point on $\cal S$

701: with a remarkable nullcline configuration.

702: 

703: Before we consider the regulation proper,

704: let us examine how the covariances

705: change across the $(w\ee ,w\ie )$ plane.

706: Figure 3b shows the values of $\bar c\ee $,

707: the time average of the instantaneous variance of $s(t)$,\footnote

708: {This corresponds, in the original system,

709: to the population- {\em and} time-average of the covariance,

710: $\langle \bar c_{ij}\ee \rangle$;

711: the latter becomes $\bar c\ee $

712: in the thermodynamic limit $N\rightarrow\infty$.

713: In the regulation equation,

714: we use the {\em instantaneous} covariance $c\ee (t)$

715: rather than its time average $\bar c\ee $ (see Discussion).

716: The time-averaged variance $\bar c\ee $

717: is used here for illustration purposes only.

718: In order to obtain an essentially constant value for $\bar c\ee $

719: rather than an oscillating function of time,

720: different values of $\rho$ are used for the two averaging operations:

721: the kernel used to average $c\ee $ into $\bar c\ee $

722: is ten times broader than the kernel used to compute $\bar s$ from $s$.}

723: along several horizontal lines in the $(w\ee ,w\ie )$ plane.

724: As expected, $\bar c\ee $ is positive only in region $\cal P$,

725: where the dynamics is periodic;\footnote

726: {In general, positive average covariance across a neuronal population

727: indicates collective fluctuations;

728: in our simplified two-dimensional system,

729: the only possible nontrivial asymptotic behavior

730: is periodic oscillation.}

731: although not shown, the same is true of $\bar c\ie $,

732: the time average of the E-to-I covariance.

733: Note that as $w\ee $ crosses the $\cal O$-to-$\cal P$ boundary

734: (Hopf bifurcation) from left to right,

735: $\bar c\ee $ increases {\em smoothly} from 0 to positive values:

736: as discussed above, the amplitude of the limit cycle

737: at this bifurcation is infinitesimal.

738: In contrast, the change in $\bar c\ee $ and in $\bar c\ie $

739: at $\cal S$ (saddlenode bifurcation) is a sharp one,

740: as the system undergoes there a transition from a {\em large} limit-cycle regime

741: to a fixed-point attractor.

742: 

743: We now start our study of covariance plasticity

744: by regulating parameter $w\ee $ in system \ref{sys_red}

745: while all other parameters, including $w\ie $, remain fixed.

746: The system under study then consists

747: of coupled equations \ref{sys_red}, \ref{covee}, \ref{regee}.

748: Equation \ref{regee} prescribes an increase of $w\ee $ when $c\ee > \theta\ee $,

749: and a decrease when $c\ee < \theta\ee $.

750: Referring to Figure 3b, we see that to the left of $\cal S$,

751: where $c\ee $ is high, the first of the two conditions applies;

752: in this region $w\ee $ increases.

753: To the right of $\cal S$ the covariance vanishes, and $w\ee $ decreases.

754: Therefore, $w\ee (t)$ is attracted to the transition line $\cal S$.\footnote

755: {The control parameter $\theta\ee $ should be smaller

756: than the value of $\bar c\ee $ immediately to the left of $\cal S$.

757: The portion of the boundary line

758: where the bifurcation is a subcritical Hopf rather than a saddlenode

759: (footnote 5) yields similar behavior,

760: since the disruption of the large-amplitude limit cycle

761: occurs very near the emergence of point attractors (see also footnote 11).}

762: \vglue .4cm

763: \centerline{\em (Insert Figure 4 around here)}

764: \vglue .4cm

765: The behavior of this $w\ee $ regulation loop is illustrated in Figure 4a

766: for the following setting of parameters:

767: $w\ei = 10$, $w\ii = 6$, $\rho=.1$, $\theta\ee =.01$, $\varepsilon\ee =.01$.

768: This figure focuses on a small region of the $(w\ee ,w\ie )$ plane,

769: and shows the projection of the trajectory of $(s, \sigma, w\ee ,w\ie )$.

770: Several trajectories are shown, all horizontal since $w\ie $ is a constant,

771: These trajectories terminate on the critical line $\cal S$,

772: and the behavior of the $s$ and $\sigma$ components on them is as follows.

773: On the trajectories coming from the left, in the $\cal P$ region,

774: $(s,\sigma)$ moves along a cyclic orbit,

775: whose amplitude grows as $w\ee $ increases

776: and approaches the bifurcation line.

777: On the trajectories coming from the right, in the $\cal T$ region,

778: $(s, \sigma)$ stays in one of the two point attractors while $w\ee $ decreases

779: until it reaches the bifurcation curve.

780: When $\cal S$ is reached, either from the left or from the right,

781: motion does not really stop.

782: Rather, $w\ee $ sets in a periodic oscillation of small amplitude

783: synchronized with a large-amplitude periodic motion of $(s,\sigma)$;

784: the frequency of this oscillation 

785: is several orders of magnitude slower than in $\cal P$,

786: hence covariance is small---it

787: matches, on average, the control parameter $\theta\ee $.

788: When in this regime, the system spends a long time in one of the two 

789: almost-attracting corners of the $[-.5,+.5]^2$ box before leaving it

790: and moving rapidly to the other corner.

791: This results in an almost-square wave,

792: a behavior that is intermediate between the fast periodic motion observed

793: in $\cal P$ and the bistable situation prevailing in $\cal T$.

794: The period of this oscillation and the amplitude

795: of the oscillation of $w\ee $

796: depend on parameters $\rho$, $\theta\ee $, and $\varepsilon\ee $.\footnote

797: {Not shown on Figure 4a is the leftmost part of region $\cal P$,

798: near the Hopf bifurcation,

799: where the limit cycle is of small amplitude

800: hence the condition $\bar c\ee  > \theta\ee $ is not realized.

801: When initialized there,

802: the system does not converge to $\cal S$.

803: However, in both the $w\ee $ and the $w\ie $ directions,

804: the domain of attraction of $\cal S$ extends to $+\infty$.}

805: 

806: We next consider the $w\ie $-regulated system,

807: where $w\ee $ and all other parameters remain fixed.

808: This system consists of coupled equations \ref{sys_red}, \ref{covie}, \ref{regie}.

809: As noted, the E-to-I covariance $c\ie $ vanishes

810: outside region $\cal P$, just like $c\ee $;

811: within $\cal P$ it varies, in a first approximation, like $c\ee $.

812: Since we chose $\varepsilon\ie $ to be {\em negative,}

813: $w\ie $ decreases in $\cal P$ and increases in $\cal T$,

814: whereas the opposite was true of $w\ee $

815: when it was regulated.

816: Figure 4b shows this $w\ie $ dynamics

817: in the same region of the $(w\ee ,w\ie )$ plane as before.

818: Parameters are $w\ei = 10$, $w\ii = 6$, $\rho=.1$,

819: $\theta\ie =.01$ and $\varepsilon\ie =-.01$.

820: The trajectories are now parallel to the $w\ie $ axis,

821: and $(w\ee ,w\ie )$ is again attracted to the critical line $\cal S$

822: separating region $\cal P$ from region $\cal T$.

823: This is true only to the left of the vertical asymptote of that curve;

824: trajectories to the right of that line go to $+\infty$.

825: 

826: In sum, regulation of either one of the two parameters $w\ee $, $w\ie $

827: has the effect of bringing the system

828: to the critical surface $\cal S$ separating the region of oscillation

829: from the region of bistable steady firing;

830: the nullcline diagram is then as in Figure 2b.

831: Note that when the system is on $\cal S$,

832: a small perturbation in the weights will elicit either oscillation,

833: constant firing at near-maximum rate,

834: or constant firing at near-minimum rate.

835: 

836: We now turn to the behavior of the system

837: when the two regulation loops act simultaneously;

838: we thus study the system of coupled equations

839: \ref{sys_red}, \ref{covee}, \ref{regee}, \ref{covie}, \ref{regie}.

840: Figure 4c shows the $(w\ee ,w\ie )$ dynamics for the same parameters as before,

841: i.e., $w\ei = 10$, $w\ii = 6$, $\rho=.1$, $\theta\ee =.01$, $\varepsilon\ee =.01$,

842: $\theta\ie =.01$ and $\varepsilon\ie =-.01$.

843: It appears from this diagram that the evolution proceeds

844: in two clearly distinct stages.

845: In the first stage,

846: which could be predicted from the study of the regulation loops

847: acting separately,

848: $(w\ee ,w\ie )$ moves toward line $\cal S$.\footnote

849: {The direction of this linear motion is roughly parallel to the line $w\ee = -w\ie $.

850: This is because $\varepsilon\ee =-\varepsilon\ie $,

851: $\theta\ee = \theta\ie $,

852: and the two covariances $c\ee $ and $c\ie $ are nearly the same.

853: Another choice of parameters would result in a different slope,

854: but otherwise similar behavior.}

855: When this line is reached,

856: motion slows down considerably---typically

857: by several orders of magnitude---and

858: proceeds {\em along} the critical line,

859: eventually converging to a point on $\cal S$ denoted $G$ in Figure 4c.

860: As before, attractor $G$ is in reality a slow limit cycle,

861: of small amplitude in $w\ee $ and $w\ie $,

862: and large amplitude in $s$ and $\sigma$.

863: All four variables, $s(t)$, $\sigma(t)$, $w\ee (t)$, and $w\ie (t)$,

864: are now synchronized;

865: the distinction between slow and fast variables has thus vanished.

866: The basin of attraction of $G$ in the $(w\ee ,w\ie )$ plane

867: roughly consists of the {\em union} of the two domains of attraction

868: of $\cal S$ for the separate $w\ee $ and $w\ie $ regulation dynamics;

869: only the region to the left of and around the Hopf line

870: is not attracted to the saddlenode line $\cal S$ and eventually to $G$.

871: \vglue .4cm

872: \centerline{\em (Insert Figure 5 around here)}

873: \vglue .4cm

874: The location of $\cal S$ in the $(w\ee ,w\ie )$ plane depends

875: on the values of the fixed parameters $w\ei $ and $w\ii $.

876: The location of the attractor $G$ on $\cal S$

877: further depends on the control parameters $\theta\ee $ and $\theta\ie $.

878: When the latter are given identical values,

879: as in the case illustrated in Figure 4c,

880: the attractor $G$ has the remarkable property

881: that the $s$- and $\sigma$-nullclines stand in {\em near overlap}

882: over a large portion of the interval [-.5,+.5] (Figure 5);

883: the flow of the system in this configuration

884: nearly vanishes on a large one-dimensional manifold

885: in the two-dimensional phase space.

886: Further, $s(t)$ and $\sigma(t)$ remain nearly identical at all times.\footnote

887: {Giving different values to parameters $\theta\ee $ and $\theta\ie $

888: mostly affects the behavior of the system

889: after it has reached $\cal S$;

890: if $\theta\ee $ is larger, resp. smaller, than $\theta\ie $,

891: the state moves downward, resp. upward, on $\cal S$.

892: When $(w\ee ,w\ie )$ is on $\cal S$ but above point $G$,

893: the nullclines are tangent to each other but do not overlap;

894: such a situation is illustrated in Figure 2b.

895: When $(w\ee ,w\ie )$ is on $\cal S$ but below point $G$,

896: the nullclines do overlap, but over a smaller domain.

897: With $\theta\ee = .0118$ and $\theta\ie = .0100$,

898: the state stabilizes in the narrow three-attractor region described in footnote 5.

899: The state $(s,\sigma)$ then visits each of the three `attractors' in turn:

900: its motion consists of a succession

901: of large-amplitude oscillations (periodic attractor)

902: and of spiraling orbits around two symmetric points

903: in the interior of the large cycle (point attractors).

904: The amplitude of the motion of $(w\ee ,w\ie )$ remains small.

905: This is a mildly chaotic behavior;

906: a more pronounced chaotic behavior

907: will be described in the next section for the full system.}

908: 

909: \section{Behavior of the regulated full system}

910: \label{full}

911: 

912: Recall that system \ref{sys_red},

913: which we used so far,

914: was derived from system \ref{sys_full}

915: by eliminating the firing thresholds $h\E $ and $h\I $ (equations \ref{chvar})

916: in such a way as to make $(.5,.5)$---$(0,0)$ in system \ref{sys_red}---a

917: center of symmetry of the dynamics.

918: While easier to analyze,

919: the reduced system is less realistic.

920: There is no clear biological justification for this hard-wired symmetry;

921: moreover, when the system is in phase $\cal T$,

922: i.e., to the right of the critical surface $\cal S$,

923: it can stay for arbitrarily long periods of time in one

924: of the two fixed point attractors, e.g. in the high-activity one;

925: this is unrealistic.

926: 

927: In this section we consider a biologically more plausible way

928: of introducing symmetry in the dynamics.

929: Rather than eliminating the thresholds according to equations \ref{chvar},

930: we {\em regulate} them,

931: thereby implementing a form of `soft' symmetry.

932: Regulating the firing thresholds in a neural network

933: is a simple way to maintain the mean activity

934: around an intermediate, useful, value.

935: This may be viewed as a simplification of the regulation mechanisms

936: at work in real brains,

937: which, in all likelihood, involve systems of inhibitory neurons

938: acting on various time scales.

939: 

940: The simultaneous regulation of four parameters

941: results in a complex dynamics,

942: which makes a thorough analysis impractical.

943: We shall proceed as follows.

944: We first consider, in system \ref{sys_full},

945: the regulation of $w\ee $ and $h\E $

946: for a given setting of all other parameters.

947: We show that the system converges to the intersection of two critical curves,

948: each of which corresponds to the establishement

949: of one point of contact between the nullclines.

950: We next consider the system

951: with all four parameters $h\E $, $h\I $, $w\ee $ and $w\ie $ regulated,

952: and study the projection of the dynamics on the $(w\ee ,w\ie )$ plane.

953: There are again two stages;

954: the first essentially reproduces the behavior observed

955: with the sole $(w\ee ,h\E )$ regulation,

956: while the second is analogous to that observed

957: when regulating $w\ee $ and $w\ie $ in the reduced system;

958: this applies for a broad range of the remaining fixed parameters $w\ei $ and $w\ii $.

959: 

960: Figure 6a is the bifurcation diagram of system \ref{sys_full}

961: in the $(w\ee ,h\E )$ plane,

962: for the following values of the fixed parameters:

963: $w\ei =10$, $w\ie =10$, $w\ii =1$, $h\I =5$.

964: As before, we ignore unstable equilibria and unstable limit cycles.

965: As before there are three regions,

966: denoted respectively by $\cal O$, $\cal T$ and $\cal P$,

967: corresponding to three types of asymptotic behavior:

968: single fixed-point attractor;

969: two fixed-point attractors (high and low activity);

970: one periodic attractor.

971: We now however subdivide region $\cal O$---somewhat arbitrarily---according

972: to the location of the fixed-point attractor in the phase space:

973: the three subregions denoted ${\cal O}_h$, ${\cal O}_m$, and ${\cal O}_l$,

974: correspond, respectively, to high, middle, and low activity for this attractor.

975: The transition between region $\cal P$ and region ${\cal O}_m$

976: takes place through the familiar, smooth, Hopf bifurcation.

977: The transition between ${\cal P}$ and ${\cal O}_h$,

978: as well as its continuation between ${\cal O}_l$ and $\cal T$,

979: takes place through a saddlenode bifurcation.

980: We denote by ${\cal S}_h$ the locus of this transition;

981: it marks the appearance of a point of contact between the nullclines

982: near the high-activity corner,

983: and is thus similar to the $\cal S$ transition in the reduced system.

984: However, due to the symmetry of that system,

985: another point of contact appeared simultaneously near the low-activity corner,

986: giving rise to a double bifurcation.

987: In system \ref{sys_full} this is no more the case,

988: and the intersection of the nullclines near the low-activity corner

989: gives rise to a distinct saddlenode bifurcation line,

990: the transition between ${\cal P}$ and ${\cal O}_l$,

991: which we denote ${\cal S}_l$.

992: \vglue .4cm

993: \centerline{\em (Insert Figure 6 around here)}

994: \vglue .4cm

995: When regulating $w\ee $ according to equation \ref{regee}

996: and leaving all other parameters fixed,

997: the behavior of system \ref{sys_full} is as follows.

998: When starting in region $\cal P$ to the left of the critical line ${\cal S}_h$,

999: the system oscillates, covariance is high, hence $w\ee $ increases

1000: until it reaches the critical line ${\cal S}_h$.

1001: A point of contact is then established

1002: near the high-activity corner of the square.

1003: The system settles in a slow periodic attractor,

1004: of small amlitude in $w\ee $ and large amplitude in $(s,\sigma)$,

1005: whereby nearly all the time is spent in the high-activity state.

1006: 

1007: We now regulate the threshold $h\E $ as well,

1008: in such a way as to stabilize $\bar{s}$,

1009: the time average of $s$, around a given target value $\theta\E $:

1010: \begin{equation}

1011: \frac{dh\E (t)}{dt} = \varepsilon\E ( \bar{s}(t) - \theta\E  ).

1012: \label{reghe}

1013: \end{equation}

1014: The rate constant $\varepsilon\E$ is positive and small,

1015: and the control parameter $\theta\E $ is chosen

1016: well in the interior of the interval $(0,1)$,

1017: e.g. between .2 and .8

1018: (remember that in system \ref{sys_full}

1019: the activity variables $s$ and $\sigma$ lie in the interval $(0,1)$).

1020: To see how equation \ref{reghe} achieves the desired regulation,

1021: note for instance that, if $\bar{s}(t) > \theta\E $, $h\E $ will increase,

1022: which in turn will result in a decrease of $\bar{s}(t)$.

1023: 

1024: When both $w\ee $ and $h\E $ are regulated,

1025: the system converges to the {\em intersection}

1026: of the two critical lines ${\cal S}_h$ and ${\cal S}_l$.

1027: In effect, we saw that the full system,

1028: when at a generic point of ${\cal S}_h$,

1029: stays nearly all the time in the high-activity state;

1030: this results in a high value of $\bar {s}$.

1031: To achieve the condition $\bar {s} \approx \theta\E $,

1032: the equilibrium for equation \ref{reghe},

1033: the system can only be on ${\cal S}_l$ {\em as well.}

1034: 

1035: The joint $(w\ee ,h\E )$ dynamics is illustrated in Figure 6b,

1036: for parameters $w\ei $, $w\ie $, $w\ii $ and $h\I $ as above,

1037: and $\rho = .2$, $\varepsilon\E = .001$, $\theta\E = .5$,

1038: $\varepsilon\ee =.01$, $\theta\ee =.01$.

1039: The intersection of ${\cal S}_h$ and ${\cal S}_l$, denoted $F$ in Figure 6b,

1040: is reached from all directions in the $(w\ee ,h\E )$ plane.

1041: When coming from low $w\ee $ values,

1042: the system oscillates and converges to $F$ through region $\cal P$.

1043: When coming from high $w\ee $ values,

1044: the system reaches $F$ through region $\cal T$,

1045: where it bounces back and forth

1046: between the high- and low-activity point attractors

1047: (an oscillation much slower than in $\cal P$).

1048: 

1049: The nullcline diagram for point $F$ of Figure 6b is illustrated in Figure 6c.

1050: There are now two points of contact between the nullclines,

1051: a situation more degenerate than the one

1052: that obtains from regulating $w\ee $ only,

1053: but `equivalent' to the situation

1054: obtained in the reduced system 

1055: by regulating a single parameter, $w\ee $ or $w\ie $

1056: (compare Figure 6b to Figure 2b).

1057: What characterizes the dynamics at point $F$

1058: is that the system is on the verge of oscillation

1059: and on the boundary of each of the two steady-firing phases.

1060: 

1061: We finally consider the system

1062: with the four parameters $h\E $, $h\I $, $w\ee $ and $w\ie $ regulated.

1063: We thus include, in addition to equations

1064: \ref{sys_full}, \ref{covee}, \ref{regee}, \ref{covie}, \ref{regie} and \ref{reghe},

1065: a regulation equation for the inhibitory threshold $h\I $:

1066: \begin{equation}

1067: \frac{dh\I (t)}{dt} = \varepsilon\I ( \bar{\sigma}(t) - \theta\I  ).

1068: \label{reghi}

1069: \end{equation}

1070: As in equation \ref{reghe},

1071: the rate constant $\varepsilon\I$ is positive and small,

1072: and $\theta\I $ is chosen in the interval $(.2,.8)$,

1073: with $\theta\I \approx \theta\E $.

1074: The variables now include the activity state $(s,\sigma)$

1075: as well as the four regulated parameters

1076: $h\E $, $h\I $, $w\ee $ and $w\ie $.

1077: 

1078: Figure 7 illustrates the behavior of this system projected

1079: on the $(w\ee ,w\ie )$ plane,

1080: for the following parameter values:

1081: $w\ei = 10$, $w\ii = 6$, $\rho=.05$, $\theta\ee =.01$, $\varepsilon\ee =.01$,

1082: $\theta\ie =.01$, $\varepsilon\ie =-.005$,

1083: $\theta\E = .5$, $\varepsilon\E = .005$,

1084: $\theta\I = .5$, $\varepsilon\I = .002$.

1085: In the sequel, this parameter setting will be referred to as {\em standard.}

1086: In a first stage, the system converges to a doubly critical point $F$

1087: as described above;

1088: each such point $F$ belongs to the common boundary of the regions

1089: of oscillation, high steady firing, and low steady firing.

1090: Although we cannot thoroughly characterize the surface of $F$ points

1091: in the four-dimensional $(w\ee , w\ie , h\E , h\I )$ space

1092: as we did in the $(w\ee ,w\ie )$ plane for the reduced system,

1093: there is, as remarked above,

1094: a functional equivalence with the $\cal S$ surface.

1095: Note that the projection of the $F$ surface on the $(w\ee ,w\ie )$ plane

1096: has a shape quite similar to that of $\cal S$ in the reduced system.

1097: As before, when the system reaches a point $F$,

1098: all variables settle in a slow, synchronous, almost-periodic motion.

1099: The oscillation of $s$ and $\sigma $ is a nearly rectangular wave,

1100: the system spending nearly all its time in the two corners of the square,

1101: where the relative amount of time spent in each corner

1102: is determined according to the value of parameter $\theta\E (\approx \theta\I )$.

1103: As before too, the first stage,

1104: which consists of the convergence to a doubly critical point $F$,

1105: is robust against parameter changes;

1106: most parameters can be individually

1107: varied over several orders of magnitude

1108: without qualitatively affecting this part of the behavior.

1109: 

1110: The second stage, consisting of a much slower motion on the $F$ surface,

1111: depends on the values of the various parameters.

1112: For most parameter settings,

1113: including the standard set (see above),

1114: the behavior on this critical surface is a slow, {\em simple,} periodic motion,

1115: of large amplitude in $(s,\sigma)$

1116: and very small amplitude in $(w\ee ,w\ie )$.

1117: The system eventually settles in a periodic attractor of this simple type,

1118: denoted again $G$ in Figure 7.

1119: Figure 8a shows the $(s,\sigma)$ projection of this attractor

1120: for the standard parameter set;

1121: its $(w\ee ,w\ie )$ projection

1122: is a small cycle around point $G$,

1123: whose nullcline diagram

1124: is similar to the one shown in Figure 5 (largely overlapping nullclines).

1125: 

1126: There exists however a small region of parameter space,

1127: mostly around $\varepsilon\I \approx \varepsilon\E$,

1128: for which a variety of more complex behaviors

1129: are observed during the second stage.

1130: The following two cases are examples of such complex behavior.

1131: For parameters as above (standard) except that

1132: $\varepsilon\E = .0051$,

1133: $\varepsilon\I = .0046$,

1134: and $\theta\ee =.011$,

1135: the system settles in a complex quasi-periodic motion (Figure 8b).

1136: For parameters as standard except that $\varepsilon\I = \varepsilon\E = .005$,

1137: the system displays strongly chaotic behavior (Figure 8, c--e).\footnote

1138: {This behavior takes place only

1139: for {\em some} initial values in the $(w\ee ,w\ie )$ plane;

1140: other initial values converge to a point attractor.}

1141: Both of these behaviors are actually {\em attractors,}

1142: reached after considerable time,

1143: yet similar behaviors also take place

1144: while the system is still moving slowly on the critical surface.

1145: 

1146: To summarize, both in the reduced and in the full system,

1147: convergence to a doubly critical surface

1148: between the regions of fast oscillations

1149: and of high and low steady firing

1150: takes place reliably for a broad range of parameters.

1151: Once this doubly critical surface is reached,

1152: motion becomes slow,

1153: depends on parameters,

1154: and, when examined in detail,

1155: reveals a variety of behaviors,

1156: ranging from simple periodic firing to chaos.

1157: 

1158: \section{Discussion}

1159: \label{disc}

1160: 

1161: This paper proposes that a regulation mechanism

1162: underlies criticality in brain dynamics.

1163: In such a scheme, regulation stabilizes the dynamics near an instability.

1164: The force driving the system towards criticality

1165: is a covariance-governed modification of synaptic efficacies

1166: in a recurrent network.

1167: Although it has been argued that criticality

1168: in some physical systems may be self-organized (Bak et al. 1987),

1169: this phenomenon may not be very widespread.

1170: The nervous system is moreover regulated homeostatically

1171: to withstand perturbations of various sorts.

1172: It is then of interest to explain

1173: how criticality in brain dynamics may nevertheless arise,

1174: from a specific, well-documented, mechanism of synaptic plasticity.

1175: 

1176: The chief motivation for viewing the dynamics of the nervous sytem as critical

1177: is the observation that brains are very sensitive organs.

1178: Not only do our brains draw distinctions between stimuli

1179: that differ only in minute details,

1180: they also allow us to establish subtle yet clear-cut boundaries

1181: between cognitive categories at different levels.

1182: The various manifestations of hyperacuity in sensory systems

1183: may be no more than elaborate forms of signal amplification;

1184: however, in higher cognitive functions

1185: such as language, abstract thinking, or, for instance, artistic composition,

1186: the ability to create a new category by drawing a fine line---an

1187: ability which manifests itself early in life

1188: and stays with us for a long time---argues

1189: in favor of regulated criticality.

1190: Such a mechanism appears to be necessary

1191: in order to explain how sensitivity is {\em maintained}

1192: in the face of the profound changes

1193: that affect the connectivity of the brain

1194: throughout development and learning.

1195: If no such mechanism were present,

1196: one would expect that the ongoing modification

1197: of the networks which carry mental representations

1198: would soon bring these networks to generic states;

1199: as mentioned in Section \ref{intro},

1200: a dynamical system in a generic state does not show high susceptibility

1201: to external influences.

1202: 

1203: The emergence of new cognitive categories

1204: may in effect be likened to a process of morphogenesis in embryology,

1205: or differentiation in cell biology.

1206: A biological structure that is about to undergo differentiation

1207: is at that particular instant of time unstable,

1208: and, as the well-studied mechanism of {\em induction} shows,

1209: highly susceptible to external signals.

1210: From a dynamical-system perspective,

1211: the emergence of qualitatively new behavior,

1212: e.g., the splitting of one attractor into two, is a bifurcation.

1213: The complexification of an individual's cognitive apparatus

1214: in the course of his or her life

1215: may be viewed as an open-ended sequence of such bifurcations.

1216: Such an interpretation has been defended by Ren\'e Thom (1975),

1217: and related ideas have been expressed by several authors

1218: (e.g. van der Maas and Molenaar 1992).

1219: Thom (1975) also suggested that structurally stable non-generic singularities

1220: may arise from a process he termed the {\em stabilization of thresholds;}

1221: this process itself would result, in various biological contexts,

1222: from the reinforcment of homeostatic mechanisms.\footnote

1223: {We thank Jean Petitot for pointing out to us

1224: that regulated criticality as proposed here

1225: is closely related to Thom's ideas.}

1226: 

1227: The covariance plasticity rule we use is linear and straightforward.

1228: Equation \ref{regee} may be viewed

1229: as a mean-field version of the covariance rule

1230: used in the associative-memory literature

1231: (see e.g. Willshaw and Dayan 1990).

1232: However, we make a rather different use of this rule.

1233: In an associative-memory model,

1234: pre- and post-synaptic activities are generally assumed to be independent,

1235: yielding a zero expected value for the covariance.

1236: Weights are modified according to the {\em instantaneous} covariance,

1237: and, as noted in Dayan and Sejnowski (1993),

1238: storage is marked by the departures of the empirical average of this quantity

1239: from its expected value, which is zero.

1240: In our model,

1241: the expected covariance is positive in the oscillatory phase.

1242: The regulation mechanism acts on a {\em slow} time scale,

1243: and, although we use the instantaneous covariance

1244: in the modification rule (Equation \ref{regee}),

1245: we might as well have used the time-averaged covariance;

1246: fast variations of the instantaneous covariance are actually smoothed out

1247: in the integration of the differential equation.

1248: Of course by the very principle of regulation proposed,

1249: the system does not dwell in the oscillatory phase;

1250: in the regulated state, the average covariance is low.

1251: 

1252: The other major difference between the situation studied here

1253: and the associative-memory paradigm

1254: is the assumption of uniform weights.

1255: As noted in Section \ref{regulation},

1256: the covariance in our uniform-weight network

1257: is simply the variance of the population-averaged activity about its mean,

1258: and is always non-negative.

1259: This makes it necessary to subtract from it a positive constant $\theta\ee$

1260: in order to allow for {\em decreases} of the weights.

1261: Thus, whereas in associative-memory models a synaptic weight

1262: decreases as a result of {\em negative} instantaneous covariance

1263: between the pre- and post-synaptic neurons,

1264: the condition for weight decrease in our model

1265: is that the mean covariance be small or zero,

1266: which happens when the system is at rest in a point attractor,

1267: of either low or high activity.

1268: 

1269: The uniform-weight network used in the present study

1270: lends itself to a detailed mathematical/numerical analysis.

1271: We performed a bifurcation analysis of the continuous-time differential system

1272: that describes the behavior of this network in the thermodynamic limit.

1273: This analysis (Section \ref{full}) reveals, among other features,

1274: the existence of a critical surface ${\cal S}_h$ in parameter space,

1275: where the system undergoes an abrupt transition

1276: from oscillatory behavior to high-rate steady firing.

1277: We showed (Section \ref{full})

1278: that Hebbian modification of the E-to-E synaptic weights

1279: drives the system toward this surface ${\cal S}_h$;

1280: this is the main mechanism of regulated criticality proposed.

1281: 

1282: However, when the system is at a general position on ${\cal S}_h$,

1283: it spends most of its time in the high-activity state,

1284: which is undesirable.

1285: A more realistic situation results from introducing some form of symmetry

1286: between the high- and low-activity phases,

1287: making use of the firing thresholds.

1288: We investigated two ways---formally

1289: different but functionally equivalent---to do this.

1290: The mathematically simpler way

1291: is to enforce an accurate symmetry on the dynamics,

1292: by imposing an appropriate relationship between

1293: the firing thresholds and the synaptic weights (equations \ref{chvar}).

1294: This results in a {\em reduced} system, with only four parameters;

1295: in this system,

1296: there occurs a {\em double} bifurcation when the system traverses

1297: the critical surface---now denoted $\cal S$---separating

1298: the oscillatory phase from the bistable, high/low, steady-firing phase

1299: (Section \ref{model}).

1300: Regulation of the sole E-to-E weight brings the system

1301: to this doubly critical surface $\cal S$ (Section \ref{reduced}).

1302: 

1303: A biologically more satisfactory solution

1304: is to {\em regulate} one or both of the firing thresholds

1305: so as to control the mean firing rates (Section \ref{full}).

1306: Thus, when we regulate the threshold for the excitatory neurons

1307: in addition to the E-to-E weight,

1308: the system converges to the intersection of ${\cal S}_h$

1309: with another critical surface, ${\cal S}_l$,

1310: which separates the oscillatory phase from the low-activity fixed-point region.

1311: Intersection points between ${\cal S}_h$ and ${\cal S}_l$

1312: are again doubly critical,

1313: and they attract the system for a wide range of parameter values.

1314: 

1315: When the system is on this doubly critical surface,

1316: it takes only a small weight perturbation

1317: to induce either of the three behaviors:

1318: intrinsic oscillation (region $\cal P$ of section \ref{full}),

1319: high activity, quiescence.

1320: It is easily seen that, when in this state,

1321: the network can also be efficiently driven

1322: by a small-amplitude time-varying signal, i.e., an external field;

1323: it is thus highly sensitive to input.

1324: The stabilization at the boundary of a region of oscillatory behavior

1325: appears to be consistent with at least one conclusion

1326: that can be safely drawn from the recent literature on cortical oscillations,

1327: namely the fact that the precise conditions under which

1328: these oscillations occur are difficult to pin down.

1329: 

1330: We further investigated the effect of regulating

1331: the E-to-I weight in addition to the E-to-E weight,

1332: according to a similar covariance rule.

1333: We showed that regulating these two weights

1334: as well as the two firing thresholds results,

1335: under appropriate parametric conditions,

1336: in convergence to an even more degenerate state.

1337: When the system is in that state,

1338: its flow vanishes on an entire one-dimensional curve

1339: in the two-dimensional phase space,

1340: instead of on isolated points.

1341: This convergence is slow and parameter-dependent,

1342: yet it is interesting to note that

1343: when the system is in or near this highly degenerate state

1344: it exhibits a range of diverse behaviors,

1345: including chaos (Section \ref{full}).

1346: The chaotic behavior shown in Figure 8, c--e

1347: consists of an irregular sequence of spontaneous transitions

1348: between the three fundamental phases of the system:

1349: oscillatory, high-activity, low-activity.

1350: 

1351: While the uniform-weight network studied in this paper

1352: lends itself to a convenient mathematical analysis,

1353: it would be interesting to know whether critical behavior

1354: may arise from {\em local} covariance plasticity,

1355: where synaptic changes are made to depend on pre- and post-synaptic activities

1356: relative to individual synapses.

1357: This question should be focused by considerations

1358: about the elaborate forms of input sensitivity

1359: that could play a role in higher brain functions.

1360: 

1361: {\bf Acknowledgments} Important contributions in the early stages of this work

1362: were made by Howard Gutowitz.

1363: We thank Rob de Boer for making available to us the use of the software {\em GRIND,}

1364: which proved an extremely valuable tool in this study.

1365: Various suggestions were made by a number of colleagues,

1366: from whom we wish to thank in particular G\'erard Weisbuch, Jean Petitot,

1367: G\'erard Toulouse, Jean-Pierre Nadal, Claude Meunier, David Hansel,

1368: Ha\"{\i}m Sompolinsky, and Stuart Geman.

1369: 

1370: \vglue 1cm

1371: 

1372: \noindent{\bf References}

1373: 

1374: \begin{description}

1375: 

1376: \item Artola, A., Br\"ocher, S., and Singer, W. 1990.

1377: Different voltage-dependent thresholds for inducing long-term depression

1378: and long-term potentiation in slices of rat visual cortex.

1379: {\em Nature}, {\bf 347}, 69--72.

1380: 

1381: \item Bak, P., Tang, C., and Wiesenfeld, K. 1987.

1382: Self-Organized Criticality: An Explanation of $1/f$ Noise.

1383: {\em Phys. Rev. Lett.}, {\bf 59}, 381--384.

1384: 

1385: \item Bienenstock, E., Cooper, L.N., and Munro, P. 1982.

1386: Theory for the development of neuron selectivity:

1387: Orientation specificity and binocular interaction in visual cortex.

1388: {\em J. Neurosci.}, {\bf 2}, 32--48.

1389: 

1390: \item Borisyuk, R.M., and Kirillov, A.B. 1992.

1391: Bifurcation analysis of a neural network model.

1392: {\em Biol. Cybernetics} {\bf 66}, 319--325.

1393: 

1394: \item Dayan, P., and Willsahw, D. 1991.

1395: Optimising synaptic learning rules in linear associative memories.

1396: {\em Biol. Cybernetics} {\bf 65}, 253--265.

1397: 

1398: \item Dayan, P., and Sejnowski, T.J. 1993.

1399: The Variance of Covariance Rules for Associative Matrix Memories

1400: and Reinforcement Learning.

1401: {\em Neural Computation}, {\bf 5}, 205--209.

1402: 

1403: \item Dudek, S.M., and Bear, M.F. 1992.

1404: Homosynaptic long-term depression in area CA1 of hippocampus

1405: and effects of {\em N}-methyl-D-aspartate receptor blockade.

1406: {\em Proc. Natl. Acad. Sci. USA}, {\bf 89}, 4363--4367.

1407: 

1408: \item Glauber, R.J. 1963.

1409: Time-dependent Statistics of the Ising Model.

1410: {\em J. Math. Phys.} 4, 294--307.

1411: 

1412: \item Hebb D.O. 1949.

1413: {\em The Organization of Behavior.}

1414: Wiley, New York.

1415: 

1416: \item Fr\'egnac, Y., Shulz, D., Thorpe, S., and Bienenstock, E. 1988.

1417: A cellular analogue of visual cortical plasticity.

1418: {\em Nature}, {\bf 33}, 367--370.

1419: 

1420: \item Fr\'egnac, Y., Shulz, D., Thorpe, S., and Bienenstock, E. 1992.

1421: Cellular analogs of visual cortical epigenesis.

1422: I--Plasticity of orientation selectivity.

1423: {\em Journal of Neuroscience}, {\bf 12}, 1280--1300.

1424: 

1425: \item Guckenheimer, J., and Holmes, P. 1983

1426: {\em Nonlinear oscillations, dynamical systems,

1427: and bifurcations of vector fields.}

1428: Springer-Verlag, New York.

1429: 

1430: \item Langton, C.R. 1990.

1431: Computation at the edge of chaos:

1432: Phase transitions and emergent computation.

1433: {\em Physica D}, {\bf 42}, 12--37.

1434: 

1435: \item Linsker, R. 1986.

1436: From basic network principles to neural architecture:

1437: Emergence of spatial opponent cells.

1438: {\em Proc. Natl. Acad. Sci. USA}, {\bf 83}, 7508--7512.

1439: 

1440: \item Metzger, Y., and Lehmann, D. 1990.

1441: Learning Temporal Sequences by Local Synaptic Changes.

1442: {\em Network: Computation in Neural Systems}, {\bf 1}(3), 169--188.

1443: 

1444: \item Metzger, Y., and Lehmann, D. 1994.

1445: Learning Temporal Sequences by Excitatory Synaptic Changes Only.

1446: {\em Network: Computation in Neural Systems}, {\bf 5}, 89--99.

1447: 

1448: \item Rubin, N. 1988.

1449: Equilibrium and oscillations in stochastic neural networks.

1450: Master's thesis, The Hebrew University, Jerusalem, Israel (in Hebrew).

1451: 

1452: \item Schuster, H.G., and Wagner, P. 1990

1453: A model for neuronal oscillations in the visual cortex.

1454: 1. Mean-field theory and derivation of the phase equations.

1455: {\em Biological Cybernetics}, {\bf 64}, 77--82.

1456: 

1457: \item Sejnowski, T.J. 1977a.

1458: Storing covariance with nonlinearly interacting neurons.

1459: {\em J. Math. Biol.}, {\bf 4}, 303--321.

1460: 

1461: \item Sejnowski, T.J. 1977b.

1462: Statistical constraints on synaptic plasticity.

1463: {\em J. Theor. Biol.}, {\bf 69}, 385--389.

1464: 

1465: \item Stanton, P.K. and Sejnowski, T.J. 1989.

1466: Associative long-term depression in the hippocampus

1467: induced by hebbian covariance.

1468: {\em Nature}, {\bf 339}, 215, May 1989.

1469: 

1470: \item Thom, R. 1975.

1471: {\em Structural Stability and Morphogenesis:

1472: an Outline of a General Theory of Models}.

1473: W. A. Benjamin, Reading, Mass.

1474: 

1475: \item van der Maas, H.L.J., and Molenaar, P.C.M. 1992.

1476: Stagewise Cognitive Development: An Application of Catastrophe Theory.

1477: {\em Psychological Review}, {\bf 99}(3), 395--417.

1478: 

1479: \item Willshaw, D., and Dayan, P. 1990.

1480: Optimal Plasticity from Matrix Memories: What Goes Up Must Come Down.

1481: {\em Neural Computation}, {\bf 2}, 85--93.

1482: 

1483: \end{description}

1484: 

1485: 

1486: 

1487: \renewcommand{\baselinestretch}{1}

1488: \small

1489: 

1490: \newpage

1491: 

1492: \centerline{\psfig{file=fig1.ps,height=20cm}} 

1493: \begin{quote}

1494: Figure 1: Mean activities of excitatory and inhibitory populations

1495: in a moderate-size uniform-weight system

1496: exhibiting oscillatory behavior ($N=70$; Glauber dynamics).

1497: \end{quote}

1498: 

1499: 

1500: \newpage

1501: 

1502: \centerline{\psfig{file=fig2a.ps,height=6cm}} 

1503: \centerline{\psfig{file=fig2b.ps,height=6cm}} 

1504: \centerline{\psfig{file=fig2c.ps,height=6cm}} 

1505: \begin{quote}

1506: Figure 2: Behavior of mean-field system

1507: for different values of the E-to-E synaptic weight $w\ee $.

1508: Diagrams show trajectories and nullclines.

1509: (a) $w\ee = 12$ (parameters are as in Figure 1);

1510: all trajectories converge to a limit cycle.

1511: (b) $w\ee = \hat w\ee \sn \approx 14.22$;

1512: the system is at the saddlenode bifurcation:

1513: nullclines are tangent to each other

1514: (no trajectories shown).

1515: (c) $w\ee = 15$; nullclines intersect,

1516: the periodic attractor has vanished,

1517: two point attractors have appeared.

1518: \end{quote}

1519: 

1520: \newpage

1521: 

1522: \centerline{\psfig{file=fig3a.ps,height=8cm}}

1523: \centerline{\psfig{file=fig3b.ps,height=8cm}}

1524: \vglue 2cm

1525: \begin{quote}

1526: Figure 3: Bifurcation diagram in $(w\ee ,w\ie )$ plane.

1527: (a) Diagram shows three regions,

1528: characterizing different attractor configurations.

1529: Region $\cal O$: single point attractor, of intermediate activity level;

1530: region $\cal P$: periodic attractor, as depicted in Figure 2a;

1531: region $\cal T$: two point attractors,

1532: of high and low activity, as depicted in Figure 2c.

1533: Transitions between regions occur through bifurcations,

1534: of Hopf type between $\cal O$ and $\cal P$,

1535: of saddlenode type between $\cal P$ and $\cal T$ (curve $\cal S$),

1536: and of pitchfork type between $\cal O$ and $\cal T$.

1537: (b) Average covariance along four different lines of constant $w\ie $

1538: in the $(w\ee ,w\ie )$ plane.

1539: Note the sharp variation of the covariance on the critical line $\cal S$

1540: separating $\cal P$ from $\cal T$.

1541: \end{quote}

1542: 

1543: \newpage

1544: 

1545: \centerline{\psfig{file=fig4a.ps,height=6cm}} 

1546: \centerline{\psfig{file=fig4b.ps,height=6cm}} 

1547: \centerline{\psfig{file=fig4c.ps,height=6cm}} 

1548: \begin{quote}

1549: Figure 4: Regulation of system \ref{sys_red} by covariance plasticity.

1550: (a) $w\ee $ is regulated, $w\ie $ is constant:

1551: state converges to critical surface $\cal S$.

1552: (b) $w\ie $ is regulated, $w\ee $ is constant:

1553: state converges to critical surface $\cal S$.

1554: (c) both $w\ee $ and $w\ie $ are regulated:

1555: state converges to a particular point, denoted $G$,

1556: on critical surface $\cal S$.

1557: \end{quote}

1558: 

1559: \newpage \vglue 2cm

1560: 

1561: \centerline{\psfig{file=fig5.ps,height=16cm}} 

1562: \vglue 2cm

1563: \begin{quote}

1564: Figure 5: Nullcline diagram at point $G$ (see figure 4c).

1565: Nullclines overlap almost perfectly over much of the interval $[-.5,.5]$.

1566: \end{quote}

1567: 

1568: \newpage

1569: 

1570: \centerline{\psfig{file=fig6a.ps,height=6cm}} 

1571: \centerline{\psfig{file=fig6b.ps,height=6cm}} 

1572: \centerline{\psfig{file=fig6c.ps,height=6cm}} 

1573: \begin{quote}

1574: Figure 6: Regulation of two parameters in system \ref{sys_full}.

1575: (a) Bifurcation diagram in $(w\ee ,h\E )$ plane.

1576: (b) Regulation of $w\ee $ and $h\E $ causes convergence to point $F$,

1577: the intersection of critical lines ${\cal S}_h$ and ${\cal S}_l$.

1578: (c) Nullcline diagram at $F$.

1579: \end{quote}

1580: 

1581: \newpage \vglue 2cm

1582: 

1583: \centerline{\psfig{file=fig7.ps,height=16cm}} 

1584: \begin{quote}

1585: Figure 7: Behavior of full system under simultaneous regulation of four parameters.

1586: Diagram shows projection on $(w\ee ,w\ie )$ plane,

1587: illustrating the similarity of behavior with reduced system

1588: (compare with Figure 4c, but note difference of scales).

1589: Limits of the attraction basin to the left are roughly indicated

1590: by the starting points of the trajectories shown;

1591: attraction basin is unbounded in all other directions.

1592: \end{quote}

1593: 

1594: \newpage

1595: 

1596:   \begin{tabular}{cc}

1597:     \psfig{file=fig8a.ps,height=6cm} &  \psfig{file=fig8b.ps,height=6cm} \\

1598:     \psfig{file=fig8c.ps,height=6cm} &  \psfig{file=fig8d.ps,height=6cm} \\

1599:     \psfig{file=fig8e.ps,height=6cm} 

1600:   \end{tabular}

1601: \begin{quote}

1602: Figure 8: Various behaviors of regulated full system

1603: after it has reached critical surface (Figure 7).

1604: Diagrams show $s(t)$ for three slightly different parameter settings (see text);

1605: in all cases, the projection of the motion on the $(w\ee ,w\ie )$ plane

1606: remains of small amplitude.

1607: (a) Simple periodic attractor, point $G$ of Figure 7;

1608: similar periodic attractors are reached for most parameter settings.

1609: (b) Complex quasi-periodic attractor.

1610: (c--e) Chaotic attractor;

1611: for a given parameter setting,

1612: three diagrams corresponding to different instants of time

1613: and different lengths of time;

1614: characteristic are the irregular transitions

1615: between the high-activity, low-activity, and oscillatory phases.

1616: 

1617: \end{quote}

1618: 

1619: \end{document} 

1620: