1: %Rev. 040121A
2: %
3: %\documentstyle{article}
4: %\baselineskip=1.2\baselineskip
5: \baselineskip=1.8\baselineskip
6: \documentstyle[12pt]{article}
7: %\documentstyle[twocolumn]{article}
8: %\documentstyle[twocolumn,12pt]{article}
9: %\Large
10: \markright{}
11: \pagestyle{myheadings}
12: %
13: \topmargin=-1.0cm
14: \oddsidemargin=0truecm \evensidemargin=0truecm
15: \textheight=23cm \textwidth=16cm
16: \begin{document}
17:
18: %New Commands
19: \newcommand{\siml}{\stackrel{<}{\sim}}
20: \newcommand{\simg}{\stackrel{>}{\sim}}
21: %\baselineskip=1.2\baselineskip
22: %\baselineskip=1.8\baselineskip
23:
24:
25: %single vs. double space
26: %\renewcommand{\baselinestretch}{2}
27:
28: \noindent
29: %
30: %
31: \begin{center}
32: {\large\bf
33: Graded persisting activity of heterogeneous \\
34: neuron ensembles subject to white noises
35: %Robust activity of ensemble neurons@\\
36: %with graded persisting firings
37: %\footnote{E-print: cond-mat/0401072}
38: \footnote{E-print: q-bio.NC/0401009
39: (http://xxx.yukawa.kyoto-u.ac.jp/abs/q-bio.NC/0401009)}
40: }
41: \end{center}
42:
43: \begin{center}
44: Hideo Hasegawa
45: \footnote{Phone:+81-42-329-7482;
46: Fax: +81-42-329-7491;
47: E-mail: hasegawa@u-gakugei.ac.jp}
48: \end{center}
49:
50: \begin{center}
51: {\it Department of Physics, Tokyo Gakugei University \\
52: Koganei, Tokyo 184-8501, Japan}
53: \end{center}
54: \begin{center}
55: (Received \today)
56: %{\rm (November 10, 2000)}
57: \end{center}
58: %\maketitle
59: \thispagestyle{myheadings}
60:
61: \begin{center}
62: {\bf Abstract} \par
63: \end{center}
64:
65: \small
66:
67: Effects of distractions such as
68: noises and parameter heterogeneity have been studied
69: on the firing activity of ensemble neurons, each of which
70: is described by the extended Morris-Lecar model showing
71: the graded persisting firings with the aid of an included
72: ${\rm Ca}^{2+}$-dependent cation current.
73: Although the sustained activity of {\it single} neurons
74: is rather robust in a sense that the activity is realized
75: even in the presence of the distractions,
76: the graded frequency of sustained firings is vulnerable to them.
77: It has been shown, however, that
78: the graded persisting activity
79: of {\it ensemble} neurons
80: becomes much robust to the distractions
81: by the pooling (ensemble) effect.
82: When the coupling is introduced,
83: the synchronization of firings in ensemble neurons is
84: enhanced, which is beneficial to firings of target neurons.
85:
86: \normalsize
87:
88: \vspace{0.5cm}
89:
90: \noindent
91: {\it Keywords}: Graded persisting activity; Morris-Lecar model;
92: Pooling effects
93: \vspace{1.0cm}
94:
95: \newpage
96: \section{Introduction}
97:
98: It has been reported that persistent activity of neurons
99: is ubiquitous in living brains.
100: Such persistent, stimulus-dependent activities are expected to
101: be neuronal substrates of the short-term working memory
102: (Brody, Rome and Kepecs, 2003).
103: For example, in the prefrontal cortex of monkeys
104: which are trained by short-term tasks,
105: the persisting activity has been interpreted as a basis
106: of working memory for a few seconds
107: (Funahashi, Bruce and Goldmanrakic, 1989;
108: Miller, Erickson and Desimone, 1996;
109: Romo, Hermandez, Lemus, Zainos and Brody, 2002).
110: In the area I of goldfish hindbrain which controls the motor systems
111: deriving the eye muscle, the persisting activity is interpreted
112: as a short-term memory of the eye position that keeps the eye still
113: between saccades
114: (Pastor, Delacruz, and Baker, 1994;
115: Seung, 1996;
116: Aksay, Baker, Seung and Tank, 2000).
117: In the later example, oculomotor neurons are considered to
118: integrate the velocity information from the eye movement,
119: yielding information on the eye position.
120:
121: One of the problems of the observed persisting activity is
122: how to keep it for extended periods.
123: Most models previously proposed are based on the recurrently
124: coupled neural networks
125: (Rosen, 1972;
126: CannonCRobinson and Shamma, 1983;
127: Seung, Lee, Reis and Tank, 2000,
128: Miller, Brody, Romo and Eang, 2003).
129: These models may account for the observed property of
130: the persistent activity. They have, however, a serious drawback
131: requiring the fine tuning of the network-feedback
132: gain within a tolerance of less than 1\%\ to stabilize
133: the firing rate expressing stored information.
134: If the overall gain of the network is slightly increased
135: than the critical value, it leads to a divergent activity.
136: On the other hand, if the gain is slightly decreased than the critical value,
137: the state reduces to the stable fixed point without firings, losing
138: the memory stored in the firing rate.
139: Some homeostatic mechanisms such as an activity-dependent
140: scaling of synaptic weights (Renart, Song and Wang, 2003) have been
141: proposed for stabilizing the neural activity.
142:
143:
144: A new mechanism based on the bistability of neurons
145: has been proposed
146: (Rosen, 1972, Guigon, Dorizzi, Burnodand Schultz, 1995,
147: Koulakov, Raghavachari, Kepec and Lisman, 2002).
148: Two specific ways to realize bistable elements have
149: been discussed:
150: one by local excitation connections
151: and the other by the voltage-dependent
152: N-methyl-D-aspartate (NMDA) channel
153: (Lisman, Fellous and Wang, 1998).
154: Recurrent networks consisting of
155: bistable elements have been shown to robustly operate
156: even when the model parameters are not well tuned.
157: It is, however, unclear whether the prediction of the bistable
158: neuron is consistent with experimental data from
159: neuronal integrators.
160: For example, bistable neurons are reported to be absent
161: in the goldfish oculomotor neurons
162: which show the persistent activity
163: (Pastor {\it et al.}, 1994;
164: Aksay {\it et al.}, 2000).
165: %(Pastor, Delacruz, and Baker, 1994;
166: %Aksay, Baker, Seung and Tank, 2000).
167:
168:
169: Quite recently, the graded persistent activity has been observed
170: for {\it single} neurons in {\it vitro}
171: of the layer-V entorhinal cortex (EC) of
172: rats
173: (Egorov, Hamam, Fransein, Hasselmo and Alonso, 2002).
174: The sustained firing rate of the
175: neurons is shown to be proportional to the integral over time of its
176: previous synaptic input.
177: EC in the parahippocampal region
178: is known to be associated with the working memory
179: (Frank and Brown, 2003).
180: It has been shown that EC neurons display
181: persisting activity during the delay phase of delayed mach
182: or non-match to sample memory task (Suzuki, Muller and Desimone, 1997,
183: Young, Otto, Fox and Eichenbaum, 1997).
184: A phenomenological rate-code model including multiple
185: bistable dendrites has been proposed
186: (Goldman, Levine, Major, Tank and Seung, 2003).
187: It has been suggested that the calcium concentration is
188: a plausible candidate for the observed parametric working memory
189: (Egorov, Hamam, Fransein, Hasselmo and Alonso, 2002).
190: Several conduction-based
191: neuron models have been proposed based on ${\rm Ca}^{2+}$-ion
192: channel.
193: Lowenstein and Sompolinksy (LS) (2003)
194: have shown that
195: by the diffusion of the non-linear dynamics of
196: the calcium concentration in bistable dendrites,
197: the wave-front of calcium propagates at a speed proportional to the
198: synaptic inputs.
199: When the ${\rm Ca}^{2+}$-dependent cation is included in a neuron model
200: described by the HH-type model
201: (Shriki, Hansel and Sompolinsky, 2003), calculated firings
202: show the behavior similar to the observed one.
203: Teramae and Fukai (2003) have proposed detailed
204: conductance-based models in which both L-type Ca channel
205: and ${\rm Ca}^{2+}$-dependent cation channel are taken into account.
206:
207: It is interesting that in contrast to recurrent neural networks
208: with bistable neurons, single neurons with bistable dendrites
209: may store information in the form of the
210: graded persisting activity.
211: For a reliable performance of information storage,
212: such single neurons are required to be robust against
213: distractions such as noises and heterogeneity in neuron elements.
214: It was reported (Mainen and Sejnowsky, 1995)
215: that although firings of single neurons
216: in {\it vitro} are precise, those in {\it vivo} are quite
217: unreliable due to noisy environment.
218: This may suggest that noises in {\it vivo} might make
219: the graded persisting activity
220: of EC neurons unreliable.
221: Experiments on EC neurons {\it in vivo} are needed
222: although they are difficult owing to the problem
223: in performing the intracellular recording in awake animals.
224:
225: A small cluster of cortex generally consists of many similar neurons
226: which include some degree of heterogeneity.
227: Each neuron generates spikes which propagate to
228: synapses exciting neurons in the next stage.
229: It has been recognized that the population of neuron
230: ensembles plays important roles in the information transmission.
231: The population improves the reliability of response of noisy neurons.
232: For example, the stochastic resonance in neuron ensembles
233: is much effective than that of single neurons.
234: An advantage of the synchronized activity is a large
235: input on a target neuron.
236:
237: From these considerations, we will make, in this paper,
238: a theoretical study on the robustness of the parametric
239: working memory,
240: taking account of two factors: (1) neuron ensembles
241: with (2) the heterogeneity and noises.
242: In order to investigate
243: the property of the persisting activity
244: of an ensemble against noises and heterogeneity,
245: we will perform simulations on ensemble neurons,
246: each of which shows the persisting activity.
247: We first develop a minimum, single neuron model,
248: which is suitable for a simulation of ensemble neurons.
249: The Morris-Lecar (ML) model, which was initially proposed for a
250: barnacle giant muscle fiber, has been
251: widely employed for a study on neuron dynamics
252: (Morris and Lecar, 1981; Rinzel and Ermentrout, 1989).
253: Although the ML model is the reduced, simplified model of
254: the detailed HH model,
255: it is more realistic than the integrate-and-fire model.
256: A new variable for the calcium channel whose function
257: depends on ${\rm Ca}^{2+}$-ion concentration has been incorporated
258: to the ML model.
259: With the aid of ${\rm Ca}^{2+}$-dependent current,
260: a single ML neuron may show the graded persisting activity.
261:
262: The paper is organized as follows.
263: The property of a single (extended) ML model is discussed in Sec. 2,
264: where effects of noises and variations of model
265: parameters are investigated.
266: In Sec. 3, we have studied
267: a neuron ensemble where each neuron is described by the extended
268: ML model proposed in Sec. 2.
269: Effects of noises and
270: the heterogeneity of model parameters on
271: the persisting activity have been investigated.
272: The synchronization within neuron ensembles is also discussed.
273: Conclusions and discussions are given in Sec. V.
274:
275: %\newpage
276:
277: \section{Dynamics of a single neuron}
278: \subsection{Adopted model}
279:
280: We have adopted the extended ML neuron model
281: for a single neuron,
282: given by
283: \begin{eqnarray}
284: C \frac{d v}{dt}
285: &=&-I_{Ca} - I_{K} -I_{cat} -I_{L}
286: + a + I(t)+ \beta_v \:\xi(t), \\
287: \frac{d w_i}{dt}
288: &=& \phi \left[ \frac{w_{o}(v)-w}{\tau_{w}(v)} \right], \\
289: \frac{d z}{dt }&=& -\frac{z}{\tau_z}
290: + b + d \:[I(t) + \beta_z \:\eta(t)],
291: %\hspace{1cm} \mbox{($p=2$ to 4)}
292: \end{eqnarray}
293: where
294: \begin{eqnarray}
295: I_{Ca}&=& g_{\rm Ca} m_{o}(v)(v-v_{\rm Ca}), \\
296: I_{K}&=& g_{\rm K} \:w \:(v-v_{\rm K}),\\
297: I_{cat}&=& g_{\rm cat}\:z\:(v-v_{\rm cat}),\\
298: I_{L}&=& g_{\rm L} (v-v_{\rm L}),
299: \end{eqnarray}
300: with
301: \begin{eqnarray}
302: m_{o}(v) &=& \frac{1}{2}
303: \left[ 1+tanh \left( \frac{v-v_1}{v_2} \right) \right], \\
304: w_{o}(v) &=& \frac{1}{2}
305: \left[ 1+tanh \left( \frac{v-v_3}{v_4} \right) \right], \\
306: \tau_{w}(v) &=& sech \left( \frac{v-v_3}{2 v_4} \right).
307: \end{eqnarray}
308: The model given by Eqs. (1)-(10) takes into account
309: the non-specified cation current ($I_{cat}$)
310: besides Ca, K and leakage currents
311: ($I_{Ca}$, $I_{K}$ and $I_{L}$) which are included
312: in the original ML model
313: (Morris and Lecar, 1981; Rinzel and Ermentrout, 1989);
314: $v$ and $w$ denote the fast membrane potential
315: and the slow auxiliary variable, respectively;
316: $z$ expresses a new variable
317: for the ${\rm Ca}^{2+}$-ion concentration
318: which is relevant to the cation channel.
319: The variable $z$ controls the conductance of the cation channel
320: which yields the persistent parametric activity,
321: as will be shown later.
322: The equation of motion for $z$
323: is assumed to be given by Eq. (3), where $b$, $d$ and $\tau_z$
324: stand for the drift velocity, the coefficient, and the life time,
325: respectively.
326: Equation (3) may be justified,
327: to some extent, by the model proposed by LS
328: (Loewenstein and Sompolinsky, 2003),
329: details being given
330: in appendix A.
331: The explicit form of an input current $I(t)$ will
332: be shown shortly [Eq. (11)].
333: $\beta_v$ and $\beta_z$ express the strengths of
334: independent white noises
335: given by $\xi(t)$ and $\eta(t)$ with zero means and
336: $<\xi(t)\:\xi(t')>=<\eta(t)\:\eta(t')>=\delta(t-t')$ and
337: $<\xi(t)\:\eta(t')>=0$.
338: In this study, we have adopted the following parameters:
339: the capacitance is $C=20$,
340: the reversal potentials of
341: Ca, K, cation channels and leakage are
342: $v_{\rm Ca}=120$, $v_{\rm K}=-84$, $v_{cat}=40$
343: and $v_{\rm L}=-60$, and
344: the corresponding conductances are
345: $g_{\rm Ca}= 4$,
346: $g_{\rm K}= 8$, $g_{cat}=1$ and
347: $g_{\rm L}= 2$:
348: other parameters are $v_1=-1.2$, $v_2=18$, $v_3=12$
349: $v_4=17.4$, $\phi=0.0667$ (Rinzel and Ermentrout, 1989),
350: $a=39.6$, $b=0$, $d=0.0001$ and $\tau_z = \infty$:
351: values of $\beta_v$ and $\beta_z$ will be shown shortly
352: (in this paper, conductances are expressed in $mS/cm^2$,
353: currents in $\mu A/cm^2$, voltages in mV, the time in ms and
354: the capacitance in $\mu {\rm F}/cm^2$).
355: These parameter have not necessarily been chosen so as
356: to reproduce observed data of EC (Egorov {\it et al.,} 2002).
357:
358: \subsection{Calculated results}
359: \subsubsection{Property of the extended ML model}
360:
361: It is noted that our ML model with these parameters
362: can start firings from the zero frequency
363: when the parameter of $a$ is varied for $a_1 \leq a < a_2$
364: where the critical
365: values are $a_1=40$ and $a_2=116$ $\mu A/cm^2$
366: for $b=0$ and $I(t)=0$
367: [see Fig. 7.6 of Rinzel and Ermentrout, 1989].
368: Then our ML neuron belongs to the type-I neuron.
369: Although the linearity of the $a-f$ relation at $a \geq a_1$
370: is not perfect, it does not matter for our purpose discussing effects
371: of distractions on the persisting activity.
372:
373: We apply to our ML neuron,
374: an input signal consisting of four pulses as given by
375: \begin{equation}
376: I(t)=\sum_{n=1}^4\: I_n(t)
377: = A_i \sum_{n=1}^{4} \;
378: c_n [\Theta(t-T^{(i)}_{n})-\Theta(t-T^{(i)}_{n}-T_w)],
379: \end{equation}
380: where $A_i$ stands for the magnitude of a pulse,
381: $T^{(i)}_{n}= 1000\:n$ ms ($n$: integer), $T_w=200$ ms,
382: $c_n = 1$ ($-1$) for $n=1-3$ ($n=4$),
383: and $\Theta(t)$ is the Heaviside function.
384: We have adopted $A_i=20$ $\mu A/cm^2$ otherwise noticed.
385:
386: Equations (1)-(11) have been solved by the fourth-order Runge-Kutta
387: method with a time step of 0.01 ms, the initial condition being given by
388: $v(0)=-40$, $w(0)=0$ and $z(0)=0$.
389: Figure 1(a) shows time courses of $v$, $w$, and $z$
390: when $I(t)$ given by Eq. (11) is applied to our ML neuron,
391: $I(t)$ being depicted at the bottom of Fig. 1(a).
392: The variable $z$ is proportional to the integral on time of
393: an input signal [Eq. (3)], and it shows a step-wise behavior:
394: $z$ is changed while an input pulse given by Eq. (11) is added
395: and it stays at plateau values of $z=0.02$, 0.04 and 0.06
396: between input pulses.
397: Figure 1(b) shows time courses of currents of $I_{K}$, $I_{Ca}$
398: and $I_{cat}$. It is noted that $I_{Ca}$ and $I_{cat}$ are
399: inward (negative) currents while $I_K$ is the outward (positive) one.
400: When an input pulse is applied to a ML neuron, $z$ is increased,
401: which triggers a flow of $I_{cat}$ and the oscillation of state
402: variables. As $z$ is furthermore increased by an injected pulse,
403: the frequency of firing is increased.
404: This is more clearly seen by circles
405: in Fig. 1(c), which shows the time-dependent frequency $f(t)$ defined as
406: the inverse of the interspike interval (ISI) given by
407: \begin{equation}
408: T_{n}(t_n)=t_{n+1}-t_n,
409: \end{equation}
410: $t_n$ being the $n$th firing time defined as the time
411: when $v(t)$ crosses the threshold $\theta$ ($-10$ mV) from below.
412: Neurons fire most strongly during the transient response
413: to applied pulses and settle to a persistent firings
414: which depend on the integral on time of the past inputs.
415: An initial 200-ms depolarizing pulse results in a sustained firing of 5.1 Hz,
416: and the second and third ones increase the frequency to 7.6 and 9.0 Hz,
417: respectively, while a persisting firing backs to
418: 7.6 Hz after the fourth hyperpolarizing pulse.
419: We note that the firing frequency $f(t)$ is successively increased
420: (decreased) when depolarizing (hyperpolarizing) pulses are applied,
421: and that the frequency is constant with plateaus between applied pulses.
422: This trend is realized also in
423: histogram in Fig. 1(d) expressing the firing rate $r(t)$ for
424: various time bins of $T_b=200$ (solid curve),
425: 400 (dashed curve) and 600 (chain curve).
426: The rate with the small time bin of $T_b=200$ reproduces
427: the transient behavior while impulses are applied, although it
428: shows an oscillation at the plateau period.
429: In contrast, the rate with large $T_b=600$, the persisting
430: firings are well explained while the transient firings are
431: averaged out.
432: When the magnitude of pulses
433: $A_i$ is increased (decreased), the value of $f(t)$
434: at plateau is increased (decreased),
435: as shown by triangles ($A_i=30$) and
436: inverted triangles ($A_i=10$) in Fig. 1(c). For example,
437: the first depolarizing pulse with $A_i=$ 10, 20 and 30 $\mu A/cm^2$
438: yields the sustained firing frequencies of 2.9, 5.1 and 6.5 Hz, respectively.
439: Similarly, when the duration of pulses $T_w$ is increased (decreased),
440: the value of $f(t)$ at plateau is increased (decreased)
441: (results are not shown).
442: Calculated results are similar to those experimentally obtained for single
443: EC neurons.
444: (Egorov {\it et al.,} 2002).
445:
446: Figures 2(a) and 2(b) show the $v-w$ and $z-v$ phase planes
447: for firings shown in Fig. 1.
448: For a comparison, nullclines of Eqs. (1) and (2) for $a=39.6$,
449: $b=0$, $z=0$ and $I=0$ are shown by dashed and chain curves,
450: respectively,
451: which have a tangentially contact at $(v,w)=(-29.6, \:0.00846)$.
452: The solid curve in Fig. 2(a) shows that the cycle starts from
453: $(v,w)=(-40,\:0)$ and circles counter-clockwise
454: almost independently of the $z$ value.
455: Figure 2(b) shows that between the plateau values of
456: $z=0.02$, 0.04 and 0.06, $v$ oscillates against $z$,
457: and that $v$ changes vertically when $z$ stays at these plateau values.
458:
459: \subsubsection{Effects of model parameters of $a$ and $b$}
460:
461: We will investigate effects of parameters when their
462: values are changed.
463: Among many model parameters included in Eqs. (1)-(10),
464: we have chosen $a$ and $b$ as the parameters to be changed
465: because they are expected to play an important role
466: in stabilizing the persisting firings.
467: Figure 3(a) shows the time course of $f(t)$ for various $a$ values of
468: $a=$39, 39.6 and 41 $\mu A /cm^2$ with $b=0$.
469: For $a=41$, a neuron fires with $f=5.2$ Hz at $t < 1000$ ms
470: before the first pulse is applied, and
471: firing frequencies after the second, third and fourth inputs
472: are larger than those for $a=39.6$.
473: In contrast, the firing frequency for $a=39$
474: is smaller than that for $a=39.6$.
475:
476: Figure 3(b) shows $f(t)$ for various $b$ values
477: of $b=-2$, 0, and 2 ($\times 10^{-6}$) with $a=39.6$ $\mu A/cm^2$.
478: The result for $b=0$ is the same as that shown by circles in Fig. 1(c).
479: For $b=-2 \times 10 ^{-6}$, $f(t)$ between pulses is gradually decreased.
480: For $b=2 \times 10 ^{-6}$, in contrast,
481: the frequency shows a positive drift.
482:
483: \subsubsection{Effects of white noises}
484:
485: Next we add white noises to our ML neuron.
486: The time course of the frequency $f(t)$ when white noises
487: of $\beta_v=4$ are added
488: to the variable $v$ in Eq. (1), is shown by dots in Fig. 3(c),
489: where open circles express the result with no noises.
490: Because of added noises, the frequency of persisting
491: firings fluctuates.
492: A similar behavior has been realized
493: when noises are added to the variable $z$ in Eq. (3),
494: whose result for $\beta_z=$ 2 is shown by dots in Fig. 3(d).
495:
496: Figures 3(a)-3(d) show that a single neuron is rather robust
497: in a sense that
498: the persisting firings are possible even when distraction are added.
499: We note, however, that the frequency of the persisting firings
500: is modified by distractions.
501: The robustness is much improved
502: in ensemble neurons by the pooling effect,
503: as will be discussed in the following Se. 3.
504:
505: \section{Dynamics of neuron ensembles}
506: \subsection{Adopted model}
507: We have assumed that an ensemble consisting
508: of $N$-unit ML neurons stores information
509: from an applied signal of $I(t)$.
510: Dynamics of an adopted the extended ML neuron model is given by
511: \begin{eqnarray}
512: C \frac{d v_i}{dt}
513: &=& g_{\rm Ca} m_{o}(v_i)(v_{\rm Ca}-v_i)
514: + g_{\rm K} w_i (v_{\rm K}-v_i)
515: + g_{\rm cat}\:z_i\:(v_{\rm cat}-v_i) \\
516: &&+ g_{\rm L} (v_{\rm L}-v_i)
517: + a_i + I_i(t)+\beta_v \:\xi_i(t), \\
518: \frac{d w_i}{dt}
519: &=& \phi \left[ \frac{w_{o}(v_i)-w_i}{\tau_{w}(v_i)} \right], \\
520: \frac{d z_i}{dt }&=& -\frac{z_i}{\tau_z}
521: + b_i + d \:[I_i(t)+\beta_z \:\eta_i(t)], \\
522: I_i(t)&=& I(t)
523: + \left( \frac{J}{N-1} \right) \sum_{j (\neq i)} G(v_j),
524: \hspace{1cm} \mbox{($i=1$ to $N$)}
525: \end{eqnarray}
526: where the sigmoidal function $G(v)$
527: is given by $G(v)=1/[1+exp(v-\theta)/\alpha]$
528: with the coupling constant $J$,
529: the threshold $\theta$ ($=-10$ mV) and the width $\alpha$ ($=1$),
530: self-coupling terms being excluded.
531: $\beta_v$ and $\beta_z$ express the
532: strengths of independent white noises of $\xi_i(t)$ and $\eta_i(t)$
533: with zero means and
534: $<\xi_i(t)\:\xi_j(t')>= <\eta_i(t)\:\eta_j(t')>
535: =\delta_{ij} \delta(t-t')$ and
536: $<\xi_i(t)\:\eta_j(t')>=0$;
537: other notations are the same as in Eqs.(1)-(10).
538: The model parameters of $a_i$ and $b_i$ are assumed to
539: be heterogenous and their distributions obey the Gaussian distribution
540: with
541: \begin{eqnarray}
542: <a_i> &=& \epsilon, \\
543: <\delta a_i \: \delta a_j> &=& \lambda^2 \:\delta_{ij},\\
544: <b_i> &=& 0, \\
545: <b_i \: b_j> &=& \nu^2 \:\delta_{ij},
546: \end{eqnarray}
547: where $\delta a_i=a_i-\epsilon$.
548: We have adopted the same model parameters as for a single
549: ML model discussed in Sec. 2.1 otherwise noticed.
550:
551: \subsection{Calculated results}
552: \subsubsection{Homogeneous ensembles with no noises}
553:
554: When we perform simulations for 10-unit homogeneous
555: ML neuron with no noises and no couplings
556: ($\epsilon=39.6$, $\lambda=0$, $\nu=0$,
557: $\beta_v=0$ and $\beta_z=0$, and $J=0$), we get
558: the result in which time courses of state variables of $v$, $w$ and $z$
559: of all neurons in the ensemble
560: are the same as that shown in Fig. 1(a) for a single ML neuron;
561: $f(t)$ and $r(t)$ are also the same as those in Fig. 1(c) and 1(d).
562: In the following, we will investigate effects of the heterogeneity
563: and noises on the persisting activity in ML neuron ensembles.
564:
565: \subsubsection{Effects of heterogeneity}
566:
567: Raster in Fig. 4(a) shows firings of neuron ensembles
568: when the heterogeneity of $\lambda=1$ is introduced
569: to the parameter of $a_i$. Raster shows that
570: due to the introduced heterogeneity, firings of some neurons
571: are increased while those of some neurons are decreased,
572: compared to those in homogeneous neuron ensembles.
573: For example, five neurons with large $a_i$ fire
574: before the first impulse is applied,
575: which yields a finite firing rate at $t < 1000$ ms.
576: Nevertheless, the firing rate averaged
577: over the ensemble is not so different from that
578: of the homogeneous ensemble shown in Fig. 1(d).
579:
580: Raster in Fig. 4(b) shows firings of an ensemble
581: when the heterogeneity of $\nu=2 \:\times 10^{-6}$
582: is introduced to the parameter of $b_i$ .
583: The firing becomes irregular due to the introduced heterogeneity.
584: However, the histogram of firing rate averaged over the ensemble
585: shows a stable, persisting firing.
586:
587: \subsubsection{Effects of noises}
588:
589: We have added white noises to a ML neuron ensemble.
590: Raster of Fig. 4(c) shows firing of the ensemble when noises with
591: $\beta_v=4$ are added to the variable $v$.
592: Because of added noises,
593: firings of an ensemble become irregular.
594: Despite added noises, a stable, persisting
595: firing is realized as the averaged firing rate $r(t)$ shows,
596:
597: When white noises of $\beta_z=2$ is added to the variable $z$,
598: firings of an neuron ensemble become irregular, as
599: raster in Fig. 4(d) shows. We note from the histogram that
600: firings show a stable, persisting activity.
601:
602: \subsubsection{Effects of couplings}
603:
604: We have so far neglected the couplings among neurons ($J=0$)
605: which are now taken into account.
606: In order to examine the firing synchrony
607: in an ensemble, we consider the quantity given by
608: \begin{equation}
609: R(t)= \frac{1}{N^2} \sum_i \sum_j <[v_i(t)-v_j(t)]^2>,
610: \end{equation}
611: which is zero in the completely synchronized state.
612: By using a proper normalization factor of
613: $R_0(t)=2 (1-1/N) \gamma(t)$ which
614: expresses the $R(t)$ value for an asynchronous state, we
615: define the {\it synchronization ratio} given by
616: (Hasegawa, 2003a, 2003b)
617: \begin{equation}
618: S(t)=1-\frac{R(t)}{R_0(t)} = \frac{\zeta(t)}{\gamma(t)},
619: \end{equation}
620: where
621: \begin{eqnarray}
622: \zeta(t)&=& \frac{1}{N(N-1)} \sum_i \sum_{j (\neq i)}
623: <\delta v_i(t) \: \delta v_j(t)>, \\
624: \gamma(t)&=& \frac{1}{N} \sum_i <\delta v_i(t)^2>,
625: \end{eqnarray}
626: with $\delta v_i(t)=v_i(t)-<v_i(t)>$.
627: It is easy to see that $S(t)$ is 1 and 0
628: for the completely synchronous and asynchronous states,
629: respectively.
630:
631: The lower frame of Figs. 5(a) shows
632: the time course of the frequency $f(t)$ calculated for an ensemble
633: with the heterogeneity of $\lambda=1$
634: and no couplings ($J=0$) by 100 trials. This should be compared to
635: the histogram for the rate shown in Fig. 4(a),
636: which has been calculated by a single trial.
637: The middle frame of Fig. 5(a) expresses the synchronization
638: ratio $S(t)$, which is vanishing because
639: of the heterogeneity (with no couplings).
640: Lower frames of Figs. 5(b), 5(c) and 5(d) show $f(t)$
641: for introduced couplings of $J=20$, 40 and 60, respectively.
642: The distribution of $f(t)$ for $J=20$ become wider than that
643: for $J=0$. This may be understood as follows.
644: When any neuron with a large $a_i$ fires, its spike propagates
645: through couplings and induces firings in other neurons
646: of an ensemble. When the coupling is furthermore increased,
647: however, the distribution in $f(t)$ is reduced
648: as the lower frame of Fig. 5(d) for $J=60$ shows.
649: In contrast, the synchronization ratio $S(t)$ shown in the middle frames
650: of Fig. 5(b), (c) and (d) is gradually increased as $J$ is increased.
651: It is interesting to note that the synchronization ratio is
652: enhanced by transient impulses, and that $S(t)$ is gradually developed
653: as more impulses are applied.
654:
655: Figures 6(a)-6(d) show similar plots of $f(t)$ and $S(t)$
656: for ensembles with various couplings
657: subject to noises of $\beta_v=4$, calculated by 100 trials.
658: Lower and middle frames of Fig. 6(a) show that for $J=0$,
659: $f(t)$ has a wide distribution and $S(t)$ is zero
660: due to noises.
661: When the coupling is introduced, the distribution of $f(t)$
662: is once increased for $J=20$, but reduced for stronger
663: $J=40$ and 60 while the synchronization ratio $S(t)$
664: is monotonously increased as increasing $J$.
665: The $J$ dependence of $f(t)$ and $S(t)$ in Figs. 6(a)-6(d)
666: is the same as that in Figs. 5(a)-5(d).
667:
668:
669:
670: %\newpage
671: \section{Conclusion and Discussion}
672:
673: In this paper, we have first proposed a minimum,
674: conductance-based model
675: showing the graded persisting activity
676: by incorporating a calcium channel (Lowenstein and Sompolinsky, 2003)
677: to the ML model (Morrris and Lecar, 1981; Rinzel and Ermentrout, 1989).
678: Then, by employing the extended ML model,
679: we have studied effects of distractions such as
680: noises and the heterogeneity in model parameters
681: on the persisting activity of single and ensemble neurons.
682: The activity of single ML neurons
683: is vulnerable because the frequency
684: of persisting firings is modified by the distractions.
685: In particular, even a small $b$ ($=2 \times 10^{-6}$)
686: yields a slow drift in the sustained frequency [Fig. 3(b)].
687: This is partly due to the fact that the value of the variable $z$
688: is very small ($\sim 0-0.06$) compared to those of $v$ and $w$.
689: As discussed in appendix A, a vanishing of the drift requires $b=0$,
690: which is realized when parameters of $c_1$, $c_2$ and $c_3$
691: satisfy the condition given by $c_2=(c_1+c_3)/2$ [Eq. (A13)].
692: This condition may be relaxed when a given neuron has multidendritic
693: branches, for which $b$ is expressed by $b=<b_m>_m$, $b_m$ being
694: the $b$ term relevant to the $m$th dendrite and $<>_m$ the average
695: over multidendrites.
696: The averaging over multidendrites is expected to yield $b=0$ for
697: single neurons.
698: Even if $b$ remains finite (but small), the drift of the sustained
699: frequency may vanish in neuron ensembles [Fig. 4(b)]
700: because of the pooling (ensemble) effect which is widely observed in
701: various neuronal functions.
702: Although we have made simulations only for several typical sets of
703: parameter values,
704: they have shown that the graded persisting activity
705: of a neuron ensemble becomes more robust against distractions
706: than that of a single neuron again by the pooling effect
707:
708: The temporally correlated neuronal activity,
709: the synchrony, has been considered to be important to various neuronal
710: processing such as perception
711: (Gray and Singer 1989, Gray, K\"{o}nig, Engel and Singer, 1989)
712: and attention (Steinmetz et al. 2000,
713: Fries, Reynolds, Rorie and Desimone 2001).
714: Recently the temporal structure in neuronal activity
715: during working memory has been observed in parietal
716: cortex of monkey (Pesaran, Pezaris, Sahani, Mitra and Andersen, 2002).
717: Our calculations have shown that
718: couplings among neurons enhance the synchrony, which is
719: expected to be beneficial to firings of target neurons.
720: It is necessary to make more detailed study
721: on the interaction and synchrony in neuron
722: ensembles with the persisting activity.
723: Our calculations in this study have been based on direct
724: simulations. We are under consideration to
725: apply, to the ML model, a semi-analytical dynamical
726: mean-field theory which was successfully applied to large-scale
727: neuron ensembles described by FitzHugh-Nagumo
728: and Hodgkin-Huxley models (Hasegawa, 2003a, 2003b).
729:
730: \section*{Acknowledgements}
731: This work is partly supported by
732: a Grant-in-Aid for Scientific Research from the Japanese
733: Ministry of Education, Culture, Sports, Science and Technology.
734:
735: \newpage
736:
737: \noindent
738: {\large \bf Appendix A. Derivation of Eq.(3)}
739:
740: Assuming a long, linear dendrite for $x \in [-L, \:L]$,
741: Lowenstein and Sompolinsky (LS) (2003) have shown that
742: the ${\rm Ca}^{2+}$-dependent cation current is
743: given by
744: \begin{eqnarray}
745: I_{cat}(t)&=&g_{cat}\: z(t) \:(v-v_{cat}), \nonumber
746: \hspace{4cm}\mbox{(A1)}
747: \end{eqnarray}
748: with the spatially-summed ${\rm Ca}^{2+}$ concentration of $z(t)$
749: given by
750: \begin{eqnarray}
751: z(t)&=& \int_{-L}^{L} dx \; c(x,t),
752: \nonumber\hspace{5cm}\mbox{(A2)}
753: \end{eqnarray}
754: where $c(z,t)$ stands for the space- and time-dependent
755: ${\rm Ca}^{2+}$ concentration satisfying the
756: partial differential equation given by
757: \begin{eqnarray}
758: \frac{\partial c(x,t)}{\partial t}
759: = h(c(x,t)) + D \frac{\partial^2 c(x,t)}{\partial x^2}
760: + k(c(x,t)) \:I(t).
761: \nonumber\hspace{2cm}\mbox{(A3)}
762: \end{eqnarray}
763: Here $h(c)=-K(c-c_1)(c-_2)(c-c_3)$
764: and $k(c)=(K/2)(c_3-c_1)(c-c_1)(c-c_3)$
765: with $c_1 < c_2 < c_3$, $I(t)$ denotes an input signal,
766: $D$ the diffusion constant and $K$ the positive coefficient.
767: The functional form of $h(c)$ shows that
768: the state is bistable at $c=c_1$ and $c=c_3$
769: but unstable at $c=c_2$.
770: LS have shown that
771: the solution of Eq. (A3) for the boundary condition
772: of $c(-L,t)=u_3$ and $c(L,t)=u_1$,
773: is given by
774: \begin{eqnarray}
775: c(x,t)=c_2+ \left( \frac{c_3-c_1}{2} \right)
776: \:tanh \left[ \frac{x-P(t)}{\lambda} \right],
777: \nonumber\hspace{2cm}\mbox{(A4)}
778: \end{eqnarray}
779: with
780: \begin{eqnarray}
781: P(t)=P(0)+u t +s \int_0^t \:dt' \;I(t'),
782: \nonumber\hspace{2cm}\mbox{(A5)}
783: \end{eqnarray}
784: where
785: \begin{eqnarray}
786: \lambda&=&\left( \frac{2}{c_3-c_1} \right)
787: \sqrt{\frac{2D}{K}}, \nonumber
788: \hspace{3cm}\mbox{(A6)}\\
789: u &=& \sqrt{2DK}
790: \left[c_2-\left( \frac{c_1+c_3}{2} \right)\right], \nonumber
791: \hspace{2cm}\mbox{(A7)}\\
792: s&=&(c_3-c_1) \sqrt{2 D K}.
793: \nonumber\hspace{3cm}\mbox{(A8)}
794: \end{eqnarray}
795: Equations (A4)-(A8) imply
796: that the position of the front
797: of ${\rm Ca}^{2+}$-concentration moves with the velocity
798: of $u$, and it is proportional to
799: the integral of the input signal $I(t)$.
800: By using Eqs. (A1)-(A8), we get
801: \begin{eqnarray}
802: \frac{d z}{dt}
803: &=& \left( \frac{c_3-c_1}{2} \right)
804: \left( tanh\left[\frac{-L-P(t)}{\lambda}\right]
805: -tanh \left[ \frac{L-P(t)}{\lambda} \right] \right) \; \frac{dP(t)}{dt},
806: \nonumber\hspace{2cm}\mbox{(A9)} \\
807: &\simeq & b + d \:I(t),
808: \nonumber\hspace{10cm}\mbox{(A10)}
809: \end{eqnarray}
810: with
811: \begin{eqnarray}
812: b &=& (c_3-c_1) \:u,
813: \nonumber\hspace{2cm}\mbox{(A11)}\\
814: d &=& (c_3-c_1) \:s,
815: \nonumber\hspace{2cm}\mbox{(A12)}
816: \end{eqnarray}
817: leading to Eq. (3), where the first term is included
818: for the relaxation process of $z$.
819: Equations (A8) and (A11) show that $b$ does not vanish unless parameters
820: of $c_1$, $c_2$ and $c_3$ satisfy the condition:
821: \begin{eqnarray}
822: c_2=\frac{c_1+c_3}{2},
823: \nonumber\hspace{2cm}\mbox{(A13)}
824: \end{eqnarray}
825: although LS adopted $b=0$ assuming the condition given by Eq. (A13).
826: %we have examined the effect of $b$ in our study.
827:
828: For a single neuron
829: with $M$-unit multiple dendritic branches where the same signal is applied
830: to all synapses, the equation of motion for the averaged variable of
831: $z=\sum_{m=1}^{M} z_m \equiv <z_m>_m$
832: is again given by Eq. (A10) but with
833: $b = <b_m>_m$ and $d = <d_m>_m$ where
834: $z_m$, $b_m$ and $d_m$ denote quantities relevant
835: to the $m$th dendrite.
836: %The other advantage of multidendrite computation is to
837: %reduce the effect of noises
838: %as pointed out by LS (Loewenstein and Sompolinsky, 2003).
839:
840:
841:
842: \newpage
843: \noindent
844: {\large \bf References}
845:
846: \begin{description}
847: %\begin{references}
848:
849: \item Aksay E., Baker R., Seung H. S., and Tank D. W. (2000).
850: Anatomy and discharge properties of pre-motor neurons in the
851: Goldfish medulla that have eye-position signals during fixations.
852: {\it Journal of Neurophysiology,} {\bf 84}, 1035-1049.
853:
854: \item Brody C. D., Rome R., and Kepecs A. (2003).
855: Basic mechanisms for graded persistent activity:
856: discrete attractors, continuous attractors, and
857: dynamic representations.
858: {\it Current Opinion on Neurobiology,} {\bf 13}, 204-211;
859: related references therein.
860:
861: \item Camperi M. and Wang X. J. (1998).
862: A model of visuospatial working memory in prefrontal cortex:
863: recurrent network and cellular bistability.
864: {\it Journal of Computational Neuroscience,} {\bf 5}, 383-405.
865:
866: \item Cannon S. C.CRobinson D. A.. and Shamma S.(1983).
867: A proposed neural network for the integrator of the oculomotor systems.
868: {\it Biological Cybernetics,} {\bf 49}, 127-136.
869:
870: \item Egorov A. V., Hamam B. N., Fransein E.,
871: Hasselmo M. E., and Alonso A. A. (2002).
872: Graded persistent activity in entorhinal cortex neurons.
873: {\it Nature,} {\bf 420}, 173-178.
874:
875: \item Frank L. M. and Brown E. N. (2003).
876: Persistent activity and memory in the entorhinal cortex.
877: {\it Trends in Neurosciences,} {\bf 26}, 400-401.
878:
879: \item Fries P., Reynolds J. H., Rorie A. H. and Desimone R., (2001).
880: Modulation of oscillatory neuronal synchronization by
881: selective visual attention.
882: {\it Science,} {\bf 291}, 1560-1563.
883:
884: \item Funahashi S., Bruce C. J., and Goldmanrakic P. S. (1989).
885: Mneumonic coding of visual space in the monkeys dorsolateral cortex.
886: {\it Journal of Neurophysiology,} {\bf 61}, 331-349.
887:
888: \item Goldman M. S., Levine J. H., Major G., Tank D. W.
889: and Seung H. S. (2003).
890: Robust persistent neural activity in a model integrator with
891: multiple hysteretic dendrites per neuron. (preprint)
892:
893: \item Gray C. M., K\"{o}nig P., Engel A. K. and Singer W. (1989).
894: Oscillatory responses in cat visual cortex exhibit inter-columnar
895: synchronization which reflects global stimulus properties.
896: {\it Nature,} {\bf 338}, 334-337.
897:
898: \item Gray C. M. and Singer W. (1989).
899: Stimulus-specific neuronal oscillations in orientation
900: columns of cat visual cortex.
901: {\it Proceedings of National Academy of Sciences
902: of the United State of America,} {\bf 86}, 1698-1702.
903:
904: \item Guigon E., Dorizzi B., Burnod Y. and Schultz W. (1995).
905: Neural correlates of learning in the prefrontal cortex
906: of monkey.
907: {\it Cerebrum Cortex,} {\bf 2}, 135-147.
908:
909: \item Hasegawa H., (2003a)
910: Dynamical mean-field theory of spiking neuron ensembles:
911: Response to a spike with independent noises.
912: {\it Physical Review E,} {\bf 67}, 041903.1\\-041903.19.
913:
914: \item Hasegawa H., (2003b)
915: Dynamical mean-field theory of noisy spiking neuron ensembles:
916: Application to Hodgkin-Huxley model.
917: {\it Physical Review E,} {\bf 68} 041909.1-041909.13.
918:
919: \item Koulakov A. A., Raghavachari S., Kepec A., and Lisman J. E. (2002).
920: Model for a robust neural integrator.
921: {\it Nature Neuroscience,} {\bf 5}, 775-782.
922:
923: \item Lisman J. E., Fellous J. M., and Wang X. J. (1998).
924: A role for NMDA-receptor channels in working memory.
925: {\it Nature Neuroscience,} {\bf 1}, 273-275.
926:
927: \item Loewenstein Y. and H. Sompolinsky (2003).
928: Temporal integration by calcium dynamics in a model neuron.
929: {\it Nature Neuroscience,} {\bf 6}, 961-967
930:
931: \item Mainen Z. F., and Sejnowsky T. J. (1995).
932: Reliability of spike timing in neocortical neurons.
933: %rat cortical slice
934: {\it Science,} {\bf 268}, 1503-1506.
935:
936: \item Miller E. K., Erickson C. A., and Desimone R. (1996).
937: Neural mechanisms of visual working memory in prefrontal cortex
938: of macaque.
939: {\it The Journal of Neuroscience,} {\bf 16}, 5154-5167.
940:
941: \item Miller P., Brody C. D., Romo R. and Wang X. (2003).
942: A recurrent network model of somatosensory parametric working
943: memory in the prefrontal cortex,
944: {\it Cerebrum Cortex,} {\bf 13}, 12080-1218
945:
946: \item Morris C. and Lecar H. (1981).
947: Voltage oscillations in the barnacle giant muscle fiber.
948: {\it Biophysics} {\bf 35}, 193-213.
949:
950: \item Pastor A. M., Delacruz R. R., and Baker R. (1994).
951: Eye position and eye velocity integrators reside in separate brain-stem nuclei.
952: {\it Proceedings of National Academy of Sciences
953: of the United State of America,} {\bf 91}, 807-811.
954:
955: \item Pesaran B., Pezaris J. S., Sahani M., Mitra P. P.
956: and Andersen R. A. (2002).
957: Temporal structure in neuronal activity during working memory
958: in macaque parietal cortex.
959: {\it Nature Neuroscience,} {\bf 5}, 805-811.
960:
961: \item Renart A., Song P. and Wang X. (2003).
962: Robust spatial working memory through homeostatic synaptic
963: scaling in heterogeneous cortical networks.
964: {\it Neuron,} {\bf 38}, 473-485.
965:
966: \item Rinzel J. R. and Ermentrout G. (1989).
967: Analysis of neural excitability and oscillation.
968: In C. Koch and I. Segev, (Eds.),
969: {\it Methods in neural modeling} (pp251-291),
970: MIT press, Cambridge, MA.
971:
972: \item Romo R., Hermandez A., Lemus L., Zainos A., and Brody C. D. (2002).
973: Neuronal correlates of decision-making in secondary
974: somatosensory cortex.
975: {\it Nature Neuroscience,} {\bf 5}, 1217-1225.
976:
977: %recurent network
978: \item Rosen M. (1972).
979: A theoretical neural integrator.
980: {\it IEEE Transactions on Biomedical Engineering,} {\bf 19}, 362-367.
981:
982: \item Seung H. S. (1996).
983: How the brain keeps the eyes still.
984: {\it Proceedings of National Academy of Sciences
985: of the United State of America,} {\bf 93}, 13339-13344.
986:
987: \item Seung H. S., Lee D. D., Reis B. Y., and Tank D. W. (2000).
988: The autapse: A simple illustration of short-term analog memory
989: storage by tuned synaptic feedback.
990: {\it Journal of Computational Neuroscience,} {\bf 9}, 171-185.
991:
992: \item Shriki O., Hansel D., and Sompolinsky H. (2003).
993: Rate models for conductance-based cortical
994: neuronal networks.
995: {\it Neural Computation,} {\bf 15}, 1809-1841.
996:
997: \item Steinmetz P. N., Roy A., Fitzgerald P. J.,
998: Hsiao S. S., Johnson K. O. and Niebur E. (2000).
999: Attention modulates synchronized neuronal firing in primate
1000: somatosensory cortex.
1001: {\it Nature,} {\bf 404}, 187-190.
1002:
1003: \item Suzuki W. A., Miller E. K. and Desimone R. (1997).
1004: Object and place memory in the macaque entorhinal cortex.
1005: {\it Journal of Neurophysiology,} {\bf 78}, 1062-1081.
1006:
1007: \item Taramae J. and Fukai T. (2003), private communications.
1008:
1009: \item Young B. J., Otto T., Fox G. D. and Eichenbaum H. (1997).
1010: Memory representation within parahippocampal region.
1011: {\it The Journal of Neuroscience} {\bf 17}, 5183-5195.
1012:
1013: %\end{references}
1014: \end{description}
1015:
1016: %\noindent
1017: %Goldman M. S., Levine J. H., Major G.,
1018: %D. W. Tank, and H. S. Seung,
1019:
1020: %\end{thebibliography}
1021: \newpage
1022:
1023: \begin{figure}
1024: \caption{
1025: %Fig B.
1026: Time courses of
1027: (a) $v$, $w$, $z$ and $I$ and
1028: (b) $I_K$, $I_{cat}$ and $I_{Ca}$
1029: of a single ML neuron with
1030: $A_i=20$, $a=39.6$ and $b=0$;
1031: $w$, $z$ and $I$ are multiplied by factors of 100, 500,
1032: and 10, respectively,
1033: and are shifted downward by 100, 150, and 180, respectively;
1034: $I_{cat}$ is multiplied by a factor of 100, and
1035: $I_{Ca}$ is shifted downward by 30.
1036: (c) The time course of the frequency $f(t)$ which is the inverse of
1037: ISI, with $A_i=10$ (inverted triangles), 20 (circles)
1038: and 30 (triangles).
1039: (d) The histogram depicting the rate $r(t)$ of firings
1040: shown in (a) and (b) with the time bins of
1041: $T_b=$ 200 (the solid curve), 400 (the dashed curve)
1042: and 600 (the chain curve).
1043: }
1044: \label{fig1}
1045: \end{figure}
1046:
1047: \begin{figure}
1048: \caption{
1049: %Fig A.
1050: (a) The $v-w$ and (b) $z-v$ phase planes
1051: relevant to firings shown in Fig. 1;
1052: dashed and chain curves in (a) express nullclines
1053: of Eqs. (1) and (2)
1054: for $A_i=20$, $a=39.6$, $b=0$, $z=0$ and $I=0$.
1055: }
1056: \label{fig2}
1057: \end{figure}
1058:
1059: \begin{figure}
1060: \caption{
1061: %Fig C.
1062: Time courses of the frequency $f(t)$ of a single ML neuron
1063: (a) for $a=41$ (triangles), 39.6 (circles)
1064: and 39 (inverted triangles) with $b=0$, $\beta_v=0$ and $\beta_z=0$,
1065: (b) for $b=2 \times 10^{-6}$, 0 (circles)
1066: and $-2 \times 10^{-6}$ (inverted triangles)
1067: with $a=39.6$, $\beta_v=0$ and $\beta_z=0$,
1068: (c) for $a=39.6$, $b=0$, $\beta_v=4$, and $\beta_z=0$, and
1069: (d) for $a=39.6$, $b=0$, $\beta_v=0$, and $\beta_z=2$:
1070: an input signal is shown at bottoms, and
1071: results of (c) and (d) are calculated by ten trials,
1072: }
1073: \label{fig3}
1074: \end{figure}
1075:
1076: \begin{figure}
1077: \caption{
1078: %Fig. E.
1079: Rasters showing firings of 10-unit
1080: ML neuron ensembles and
1081: histograms expressing the firing rate $r(t)$
1082: with time bins of $T_b=$ 200 (solid curves),
1083: 400 (dashed curves) and 600 (dot-dashed curves)
1084: for
1085: (a) $\lambda=1$, $\nu=0$, $\beta_v=0$, $\beta_z=0$,
1086: (b) $\lambda=0$, $\nu=2 \times 10^{-6}$, $\beta_v=0$, $\beta_z=0$,
1087: (c) $\lambda=0$, $\nu=0$, $\beta_v=4$, $\beta_z=0$ and
1088: (d) $\lambda=0$, $\nu=0$, $\beta_v=0$ $\beta_z=2$,
1089: with $\epsilon=39.6$ and $J=0$, calculated by a single trial,
1090: an input signal being shown at bottoms.
1091: }
1092: \label{fig4}
1093: \end{figure}
1094:
1095: \begin{figure}
1096: \caption{
1097: %Fig. K.
1098: Time courses of the frequency $f(t)$ (lower frames)
1099: and the synchronization ratio $S(t)$ (middle frames)
1100: of heterogeneous ML neuron ensembles for
1101: (a) $J=0$, (b) $J=20$, (c) $J=40$ and (d) $J=60$,
1102: with $\lambda=1$, $\beta_v=0$,
1103: $\epsilon=39.6$ and $\beta_z=0$,
1104: calculated by 100 trials; an input signal is shown in upper frames, and
1105: left and right ordinates are for $f(t)$ and $S(t)$, respectively.
1106: }
1107: \label{fig5}
1108: \end{figure}
1109:
1110: \begin{figure}
1111: \caption{
1112: %Fig. K.
1113: Time courses of the frequency $f(t)$ (lower frames)
1114: and the synchronization ratio $S(t)$ (middle frames)
1115: of noisy ML neuron ensembles for
1116: (a) $J=0$, (b) $J=20$, (c) $J=40$ and (d) $J=60$,
1117: with $\beta_v=4$, $\lambda=0$,
1118: $\epsilon=39.6$ and $\beta_z=0$,
1119: calculated by 100 trials; an input signal is shown in upper frames, and
1120: left and right ordinates are for $f(t)$ and $S(t)$, respectively.
1121: }
1122: \label{fig6}
1123: \end{figure}
1124:
1125:
1126: %------------------------------------
1127: \end{document}
1128: %---------------------
1129: