1: \documentstyle[namedreferences,epsfig]{kluwer}
2:
3: \setlength{\textwidth}{6.5 in}
4: \setlength{\oddsidemargin}{0.5 in}
5: \setlength{\evensidemargin}{0.5 in}
6: \setlength{\topmargin}{-0.3 in}
7: \setlength{\textheight}{9.0 in}
8: \setlength{\headheight}{0.25 in}
9: \setlength{\headsep}{0.2 in}
10: \setlength{\footheight}{0. in}
11: \renewcommand{\baselinestretch}{1.1}
12:
13: \newcommand{\ra}{\rightarrow}
14: \newcommand{\et}{{\em et al.}}
15: \newcommand{\ie}{{i.e.}}
16:
17: \begin{opening}
18: \title{Frequency control in synchronized networks of inhibitory neurons}
19:
20: \author{Carson C. \surname{Chow}}
21: \institute{Department of Mathematics and Center for BioDynamics,
22: Boston University, Boston, MA 02215\footnote{After September 1998:
23: Dept. of Mathematics, University of Pittsburgh,
24: Pittsburgh PA 15260}}
25: \author{John A. \surname{White}}
26: \institute{Department of Biomedical Engineering and Center for BioDynamics,
27: Boston University, Boston, MA 02215}
28: \author{Jason \surname{Ritt}}
29: \author{Nancy \surname{Kopell}}
30: \institute{Department of Mathematics and Center for BioDynamics,
31: Boston University, Boston, MA 02215}
32:
33: \date{\today}
34: \end{opening}
35:
36: \begin{document}
37:
38:
39:
40: \maketitle
41:
42: \begin{abstract}
43:
44: We analyze the control of frequency for a synchronized inhibitory
45: neuronal network. The analysis is done for a reduced membrane model
46: with a biophysically-based synaptic influence. We argue that such a
47: reduced model can quantitatively capture the frequency behavior of a
48: larger class of neuronal models. We show that in different parameter
49: regimes, the network frequency depends in different ways on the
50: intrinsic and synaptic time constants. Only in one portion of the
51: parameter space, called `phasic', is the network period proportional
52: to the synaptic decay time. These results are discussed in connection
53: with previous work of the authors, which showed that for mildly
54: heterogeneous networks, the synchrony breaks down, but coherence is
55: preserved much more for systems in the phasic regime than in the other
56: regimes. These results imply that for mildly heterogeneous networks,
57: the existence of a coherent rhythm implies a linear dependence of the
58: network period on synaptic decay time, and a much weaker dependence on
59: the drive to the cells. We give experimental evidence for this
60: conclusion.
61:
62:
63:
64: \end{abstract}
65:
66: \keywords{gamma oscillations, hippocampus, interneurons, synchronization}
67:
68:
69: \section{Introduction}
70:
71: Coherent neuronal oscillations have been implicated widely as a
72: correlate of brain function~\cite{gray94,llinas93}. However, our
73: understanding of the mechanisms underlying such activity is still in
74: its infancy. (See \citeauthor{jeff96}~(\citeyear{jeff96}) and
75: \citeauthor{ritz97}~(\citeyear{ritz97}) for recent reviews). It has
76: been demonstrated
77: experimentally~\cite{whit95,jeff96},
78: computationally~\cite{wang92,whit95}, and
79: analytically~\cite{vrees94,gerstner96,terman} that mutual inhibition
80: can generate stable synchronous oscillations. Previous analytical
81: work mostly concentrated on the mechanisms responsible for
82: synchronization. Here we address the mechanisms for controlling the
83: network frequency and the interplay between frequency and synchronization.
84:
85: Previously~\cite{white97}, we showed that synchronization of
86: inhibitory networks can occur in a wide range of frequencies for
87: homogeneous neurons, but in the presence of heterogeneity
88: network synchronization is very fragile in some parameter regimes.
89: We have identified a pair of parameter regimes,
90: denoted `phasic' and `tonic'; in the former some coherence is
91: maintained in the presence of mild heterogeneity, while in the latter
92: it is lost. In this paper, we relate these regimes to the parameters that
93: determine the frequency of the network when it is synchronized.
94:
95: There are three important time
96: constants in the network dynamics. One is an intrinsic time constant of the
97: uncoupled cell, dependent on the conductances and the capacitance
98: of the membrane. The second is the decay time of the inhibition. The
99: third, the network period, depends on the other two time
100: scales and other parameters, notably the synaptic conductance and
101: injected current.
102: The aim of this paper is to understand how the first and second time
103: constants influence the third.
104:
105: We show that there are three asymptotic regimes, in which the network
106: period varies differently as the inhibitory time constant is
107: changed. Two of these are the tonic and phasic~\cite{white97} regimes
108: mentioned above and the third we call `fast'. In the tonic regime,
109: the period is small compared with both the intrinsic time scale and
110: the synaptic decay time. In this regime, we show that the network
111: frequency is only weakly dependent on the inhibitory decay time;
112: indeed, the frequency is affected mainly by the average amount of
113: inhibition, not by the time course of that inhibition. In the phasic
114: regime, the intrinsic membrane time scale is small compared with
115: the network period and the synaptic decay time. In this regime, we
116: show that the network period is proportional to the decay time, with
117: the proportionality constant a function of other network parameters.
118: In the fast regime, the synaptic decay time is short compared to
119: the network period. In this regime, the period is dominated by the
120: intrinsic time constant. We give constraints on the network
121: parameters (intrinsic membrane time scale, synaptic decay time,
122: synaptic conductance and applied current) for the network to fall into
123: each of the three asymptotic regimes.
124:
125: The three regimes can be related to the behavior of the network in the
126: presence of mild heterogeneity , as shown numerically in White
127: \et~(\citeyear{white97}). In the fast regime, the synaptic influence
128: acts as a brief inhibitory pulse which has been shown to lead to
129: stable anti-synchronous oscillations
130: \cite{friesen94,perkel74,skinner94,vrees94,wang92}. In the phasic and
131: tonic regimes, synchrony through inhibition is fairly robust for
132: homogeneous networks of neurons over a wide range of network
133: frequencies~\cite{vrees94,gerstner96}. However,
134: it is very fragile when even mild heterogeneity is included
135: \cite{golomb93,wang96,white97}. We showed that the loss of coherence
136: of the network happens by different mechanisms in the tonic and phasic
137: regimes \cite{white97}. In the tonic regime, mildly heterogeneous
138: networks are effectively de-coupled and exhibit asynchrony (loss of
139: phase relationships among the cells); in the phasic regime, the
140: network can lose some coherence via suppression, in which the slower cells
141: receive enough inhibition to prevent their firing.
142:
143: Our analytical work uses a reduced membrane description for the neuron
144: with a biophysically-based synaptic model. The analysis is for a
145: homogeneous network; when considering the frequency of a synchronous
146: solution, we can regard the network as a single self-inhibited neuron.
147: We show that in the phasic regime, the synaptic current dominates and
148: the actual intrinsic membrane dynamics are not as important for the
149: firing frequency. In the tonic regime, the intrinsic dynamics do
150: become more important. We give a comparison between our
151: analytical estimates for the period and those obtained numerically
152: for conductance-based neuron models.
153:
154: The work in this paper, in conjunction
155: with~\citeauthor{white97}~(\citeyear{white97}) shows that, in
156: the presence of mild heterogeneity, the control of frequency is
157: strongly tied to the time scale of the mechanism that produces the
158: synchronization. That is, in order to have coherence for a mildly
159: heterogeneous system, the latter must be in the phasic regime, which
160: then implies that the frequency is controlled by the time constant of
161: the inhibitory decay. It was found experimentally and numerically
162: that the frequency of coherently firing interneurons in the CA1 region
163: of the hippocampus is strongly dependent on the decay time of the
164: inhibitory synapse~\cite{whit95,jeff96,traub96,wang96}. The analytical
165: work given in this paper clarifies the reasons for this frequency
166: behavior as well as the remark found in many papers that the
167: inhibitory decay time can be a critical factor in the determination of
168: the network frequency~\cite{destexhe93,skinner93,kopell}.
169:
170:
171:
172: \section{Neuron model}
173:
174: We consider the frequency control of so called Type I
175: neuronal dynamics~\cite{hansel95,bard96} and examine
176: simplifications that would make it more amenable to
177: analysis. These neurons are distinguished by the fact
178: that they have positive {\em phase response
179: curves} (PRC)~\cite{hansel95}. \ie~a positive depolarizing current always
180: advances the time of the next spike.
181: It has been shown recently that neurons which
182: admit very low frequency oscillations near the critical applied
183: current are Type I~\cite{bard96} although the converse is not
184: necessarily true. We note that the type of bifurcation to firing of
185: the neuron is not important for our
186: analysis. Type I neurons
187: have been used to represent inhibitory interneurons
188: in the hippocampus~\cite{traub96,wang96,white97}.
189:
190: We consider an inhibitory network of single-compartment neurons
191: with membrane
192: potential dynamics of the form~\cite{rinzel,bard96,hansel95,wang96,white97}
193: \begin{equation}
194: C\frac{dV_i}{d\tilde{t}}= \tilde{I}_i- \sum I_{\rm ion} - I_s,
195: \label{mem1}
196: \end{equation}
197: where $\tilde{I}_{i}$ is an applied current, $I_{\rm ion}$ are the
198: ionic currents responsible for spike generation and recovery and $I_s$
199: is the synaptic current induced by the spikes of other neurons coupled
200: pre-synaptically. The ionic currents $I_{\rm ion}$ are functions of
201: potential and the dependent dynamical variables, which are in turn
202: governed by a system of differential equations. In the Appendix we
203: give the equations for two conductance-based neuron
204: models~\cite{white97,ermenkopell98}, which are Type I or at least very
205: close to Type I (data not
206: shown). Figure~\ref{fig:spike} shows example voltage traces for the
207: two models. The coupling between the neurons is exclusively through
208: chemical synapses represented by the synaptic current which takes the
209: form $I_s=\tilde{g} S(\tilde{t})(V-V_s)$, where
210: \begin{equation}
211: \frac{dS}{d\tilde{t}} = \tilde{\alpha} F(V_{pre})(1-S)-\tilde{\beta} S,
212: \label{synapse}
213: \end{equation}
214: $\tilde{\alpha}$ and $\tilde{\beta}$ are respectively the
215: synaptic rise and decay rates, $V_{pre}$ is the membrane potential of
216: the pre-synaptic neuron,
217: and $F(V)=1/(1+\exp[-V])$. We will
218: often consider the synaptic decay time
219: $\tilde{\tau}\equiv\tilde{\beta}^{-1}$.
220:
221: We approximate the full conductance-based dynamics with
222: integrate-and-fire dynamics where the firing of the neuron is
223: represented by the resetting of the membrane potential whenever it
224: crosses a threshold. Our justification for using this
225: approximation hinges on three considerations: 1) both
226: the conductance-based and integrate-and-fire models
227: are Type I (in the sense of positive PRC), 2) the action potentials
228: (spike widths) are narrow compared to their typical spiking period so
229: the frequency is dominated by the membrane recovery time, and 3) the time
230: scale for spike generation is very fast compared to the recovery time
231: so an effective threshold for spiking can be defined. The first point
232: was verified by observing that the measured PRC of the
233: conductance-based models near
234: the bifurcation to firing is positive.
235:
236: We model the approach to threshold with a simple
237: passive decay to obtain dynamics governed by a single equation of the
238: form:
239: \begin{equation}
240: C\frac{dV}{d\tilde{t}}\simeq \tilde{I} - g_m(V-V_r) - \tilde{g}
241: S(\tilde{t})(V-V_s),
242: \label{if1}
243: \end{equation}
244: where $g_m$ is an effective membrane recovery conductance, $V_r$ is an
245: effective membrane reversal potential and $V_s$ is a synaptic reversal
246: potential. $V(t)$ is reset to $V_0$
247: whenever it reaches the threshold potential $V_T$.
248: The passive decay to threshold is a very good approximation for some
249: neuronal models such as the reduced Traub and Miles
250: model given in the Appendix~\cite{ermenkopell98}. On the other hand,
251: we will show that even
252: when the passive decay is not a good approximation to the slow
253: dynamics of the neuron model, it can still adequately describe the
254: frequency behavior, especially in the phasic regime where the synaptic
255: current dominates.
256:
257: The synaptic current $S(\tilde{t})$ is generated from the spikes of
258: pre-synaptic neurons. This
259: must be emulated in the reduced model (\ref{if1}). Here we consider
260: $S(\tilde{t})$ to be an arbitrary time dependent function. In
261: Sec.~\ref{sec:syn}, we analyze some biophysical synaptic models in detail and
262: explicitly derive the time course of $S(\tilde{t})$ in response to a
263: pre-synaptic spike.
264:
265: To simplify the analysis, we rescale Eq.~(\ref{if1})
266: so that only dimensionless parameters remain.
267: The voltage can be rescaled via
268: $v = (V-V_0)/(V_T-V_0)$, so that
269: the reset potential is at $v=0$ and the threshold is at $v=1$. We
270: define an intrinsic membrane decay time
271: \begin{equation}
272: \tau_m\equiv C/g_m
273: \end{equation}
274: and rescale time by $t=\tilde{t}/\tau_m$.
275: This scaling takes the membrane time scale to 1,
276: leading to the dimensionless equation
277: \begin{equation}
278: \frac{dv}{dt}= I - [1 + \gamma S(t)]v - g S(t),
279: \label{re}
280: \end{equation}
281: where
282: \begin{equation}
283: I=\frac{\tilde{I}+ I_r}{I_T}, \qquad g= \frac{\tilde{g}}{g_T},
284: \qquad \gamma = \frac{\tilde{g}}{g_m},
285: \label{transform}
286: \end{equation}
287: with
288: $I_T=g_m (V_T-V_0)$, $I_r= g_m(V_r-V_0)$, and $g_T = I_T/(V_0-V_s)$.
289: The neuron is said to fire
290: each time $v(t)$ reaches 1, whereupon it immediately resets to
291: $v=0$. It is important to note that this does not restrict
292: $v(t)$ to nonnegative values away from $t=0$.
293:
294: An even simpler model is
295: \begin{equation}
296: \frac{dv}{dt}=I - v - g S(t).
297: \label{re2}
298: \end{equation}
299: which follows from Eq. (\ref{re}) if $\gamma << 1$.
300: We refer to the reduced model given by Eq.~(\ref{re}) as model 1 and
301: the one given by Eq.~(\ref{re2}) as model 2.
302: The two models differ in
303: that the synapse acts solely as a forcing function in model 2, but
304: also affects the membrane decay rate in model 1.
305: As we will show, the frequency behavior of the two models is
306: qualitatively
307: similar
308: even if $\gamma$ is not small.
309: We will focus most attention on model 2 in our analysis. The
310: synchronization tendencies of model 2 for slow synapses has been studied
311: thoroughly \cite{vrees94,hansel95,gerstner96,chow97}.
312:
313: \section{Synaptic Model}\label{sec:syn}
314:
315: In conductance-based neuron models, the post-synaptic current is
316: initiated by
317: a pre-synaptic spike. In the reduced model, the spikes have been
318: eliminated so this process must be modeled. We do so by
319: deriving the synaptic time course for Eq.~(\ref{synapse})
320: explicitly for a stereotypical pre-synaptic spike.
321: As we will show below, the spike response can be
322: described by a time dependent recursive forcing function of the form
323: \begin{equation}
324: S(t) \ra a(t-t_l) S(t) + S_f(t-t_l),
325: \label{recur}
326: \end{equation}
327: where $S(t)$ is the synaptic gating variable to be used in
328: Eq.~(\ref{re}), $t_l$ gives the spiking times of the
329: pre-synaptic neuron, $S_f(t)$ is the stereotypical post-synaptic
330: response, and $a(t)$ is a `memory' function. (Note that all time and
331: rate parameters
332: have been rescaled by $\tau_m$ in this section).
333: We divide synapses
334: into two types -- {\em saturating} and {\em nonsaturating}. If the
335: synaptic variable $S(t)$ increases without bound as the rate of firing
336: of the pre-synaptic cell approaches infinity then we call this a
337: nonsaturating synapse. If however, $S(t)$ saturates to a fixed value
338: as the firing rate approaches infinity then we call this a saturating
339: synapse. The behavior of the function $a(t)$ determines the type of
340: the synapse.
341:
342: We first derive the recursion relation explicitly for
343: a simple nonsaturating synaptic model given by
344: \begin{equation}
345: \frac{dS}{dt}= -S/\tau + \sum_{l=0}^{N} \delta(t-t_l),
346: \label{synns}
347: \end{equation}
348: where $t_l$ denotes the times of the pre-synaptic spikes.
349: Synapses of this form have been considered in
350: previous models~\cite{abbott93,tsodyks93,vrees94,hansel95}. Integrating
351: Eq.~(\ref{synns}) results in
352: \begin{equation}
353: S(t) = S(0)e^{-t/\tau} + \sum_{l=0}^N e^{-(t-t_l)/\tau},
354: \end{equation}
355: where we begin the integration at $t=0$.
356: This sum is equivalent to the recursion relation
357: \begin{equation}
358: S(t) \ra S(t) + e^{-(t-t_l)/\tau}.
359: \label{vvrec}
360: \end{equation}
361: Here $a(t)=1$ and $S(t)$ increases without bound as the frequency
362: increases.
363:
364:
365: Now consider the synaptic
366: model in the Appendix. We will show that this is a saturating
367: synapse. When the
368: pre-synaptic neuron fires, $F(V_{\rm pre})$ in Eq.~(\ref{synapse})
369: rises from near zero to a value near unity, then returns to zero. The
370: precise shape is determined by the temporal characteristics of the
371: action potential and $F(V)$. Our analysis is similar to that
372: of~\citeauthor{destexhe94}~(\citeyear{destexhe94}).
373: We assume that the shape of $F(V)$ can be
374: approximated by a square pulse with a width of $\Delta t$, given by
375: the time the membrane potential remains above zero during an action
376: potential (\ie~width of the spike). We consider a train of square
377: pulses arriving at times
378: $t_l$, with $t_l + \Delta t < t_{l+1}$ \ $\forall\,l$ (\ie~no
379: overlapping spikes). The synaptic response is then found from
380: \begin{equation}
381: \frac{dS}{dt}=\alpha Q(t)(1-S)-\beta S,
382: \label{syn2}
383: \end{equation}
384: \begin{equation}
385: Q(t) = \sum_{l=-\infty}^{\infty} H(t-t_l),
386: \end{equation}
387: where $H(t)$ is a square pulse of unit height with rise at $t=0$ and fall
388: at $t= \Delta t$.
389:
390: Given $l$, we define
391: \begin{equation}
392: R(t) \equiv \alpha \int_{t_l}^{t} Q(t') dt'
393: = \left\{\begin{array}{ll}
394: \alpha (t-t_l), & t_l \le t \le t_l+\Delta t\\
395: \alpha \Delta t, & t_l + \Delta t < t < t_{l+1}
396: \end{array}
397: \right. \end{equation} and then integrate~(\ref{syn2}) from
398: $t_l$ to $t<t_{l+1}$ to yield \begin{eqnarray} S(t) &=&
399: e^{-R(t)-\beta t} \left[S(t_l) e^{\beta t_l} +
400: \int_{t_l}^t \alpha Q(t') e^{R(t')+\beta t'}
401: dt'\right] \\ &=& \left\{ \begin{array}{ll}
402: e^{-\alpha(t-t_l)}(S(t_l)e^{- \beta(t-t_l)}) +
403: \left(\frac{\alpha}{\alpha+\beta}
404: \right)\left(1-e^{-(\alpha + \beta)(t-t_l)}\right), &
405: t_l \le t \le t_l+\Delta t \\ e^{-\alpha \Delta t} (S(t_l) e^{
406: -\beta(t-t_l)}) + \left(
407: \frac{\alpha}{\alpha+\beta}\right) \left(1 -
408: e^{- (\alpha+\beta) \Delta t}\right)e^{-
409: \beta(t-t_l-\Delta t)}, & t_l + \Delta t < t < t_{l+1}
410: \end{array}\right. \label{S} \end{eqnarray}
411: Consider $S(t)$ at the time of the next spike, $t=t_{l+1}$. If the
412: spike does not occur, $S(t)$ will continue to decay exponentially with
413: rate $\beta$, \ie
414: \begin{equation}
415: \begin{array}{ll}
416: S(t)=S(t_{l+1})\exp(-\beta(t-t_{l+1})), & t_{l+1} < t.
417: \end{array}
418: \label{nospike}
419: \end{equation}
420: We can thus rewrite
421: Eq. (\ref{S}) as a recursion relationship with the spike at $t_{l+1}$ as
422: \begin{equation}
423: S(t)\ra a(t-t_{l+1}) S(t)+S_f(t-t_{l+1})
424: \label{rec}
425: \end{equation}
426: where
427: \begin{equation}
428: a(t) = \left\{\begin{array}{ll}
429: e^{-\alpha t},
430: & 0\le t\le \Delta t,\\
431: e^{-\alpha \Delta t},
432: & \Delta t < t,
433: \end{array} \right.
434: \label{f}
435: \end{equation}
436: and
437: \begin{equation}
438: S_f(t) = \left\{\begin{array}{ll}
439: (\frac{\alpha}{\alpha+\beta})
440: (1-e^{-(\alpha+\beta)t}),
441: & 0\le t\le \Delta t, \\
442: (\frac{\alpha}{\alpha+\beta})
443: \left(1 -
444: e^{-(\alpha+\beta) \Delta t}\right)e^{-\beta
445: (t-\Delta t)},
446: & \Delta t < t.
447: \end{array} \right.
448: \label{Sf}
449: \end{equation}
450: The $S(t)$ appearing in the right hand side of Eq. (\ref{rec}) is
451: understood to be the function that would occur in the absence of a new
452: spike, Eq. (\ref{nospike}).
453:
454: In contrast to Eq. (\ref{vvrec}), there is a multiplicative
455: factor $a(t)$ which `damps' away
456: the past. For $a(t)<1$, $S(t)$ always saturates to a finite value.
457: $S_f(t)$ and $a(t)$ have discontinuous first derivatives because a
458: square pulse was used as input, but in general the functions will be
459: smooth.
460:
461: Depending on the
462: values of the three parameters $\alpha$, $\beta$ and
463: $\Delta t$, the synaptic time course can take on many shapes. Here we
464: are concerned with narrow spikes (small $\Delta t$) and slowly
465: decaying inhibition ($\beta << \alpha$).
466: If the rise time is very fast, we can explicitly take the double limit
467: $\alpha\ra\infty$, $\Delta t \ra0$, $\alpha\Delta t=c$,
468: where c is a constant, to get
469: \begin{equation}
470: S_f(t)= (1-e^{-c})e^{-\beta t}.
471: \label{histupdate}
472: \end{equation}
473: and $a=e^{-c}$.
474:
475: The constant $c$ is the ratio of the spike width to the rise time of
476: the synaptic current, and determines the contribution from the past to
477: the current synaptic response. If $c$ is large enough we can
478: reasonably ignore the past and use the approximation $a\simeq 0$ and
479: $S_f(t)\simeq\exp(-\beta t)$. Then the synaptic update
480: function to be used in Eqs.~(\ref{re}) and (\ref{re2}) takes the form
481: \begin{equation}
482: S(\tilde{t})\rightarrow e^{-(t-t_l)/\tau},
483: \label{synupdate}
484: \end{equation}
485: where
486: $\tau=\tilde{\tau}/\tau_m$ is the rescaled synaptic decay time.
487: Equation (\ref{synupdate}) gives the recursion relation for the synapse
488: where $a(t)=0$ and
489: is valid in the limit where the rise time
490: of the synaptic gating variable is very fast compared to the spike
491: width. We note that a nonzero rise time is very important for
492: synchronization~\cite{vrees94,hansel95,gerstner96,terman,chow97} but
493: it is not as important for frequency determination.
494:
495:
496:
497:
498: \section{Period of a synchronized network}\label{freq}
499:
500: In this section we will derive the form of the network period
501: analytically for model 2 [Eq.~(\ref{re2})] in
502: terms of a transcendental relation for saturating and non-saturating
503: synapses.
504: We will also analyze the differences that would
505: result if we considered model 1 [Eq.~(\ref{re})]. In general the same
506: qualitative behavior holds
507: between the two reduced models.
508: To calculate the period for the reduced models, we assume
509: the neuron spikes repeatedly with a period $T$ (the
510: frequency is defined as $f=T^{-1}$). We then
511: integrate the membrane equation from the time of the last spike to the
512: time of the next spike. The period is obtained self consistently by
513: imposing the constraint $v(T)=1$.
514:
515: Suppose the neuron last fired at $t=0$, then
516: for model 1 the membrane voltage obeys
517: \begin{equation}
518: v(T)=1=e^{-\mu(T)}\int_0^T e^{\mu(t')}[I- g S(t')] dt'
519: \label{v2}
520: \end{equation}
521: where (for model 1)
522: \begin{equation}
523: \mu(t)=\int_0^t[1 + \gamma S(t')] dt'.
524: \label{mug}
525: \end{equation}
526: For saturating synapses, $S(t)= \exp(-t/\tau)$, which implies that
527: \begin{equation}
528: \mu(t)= t + \gamma (1-e^{-t/\tau})\tau.
529: \label{mu}
530: \end{equation}
531: For model 2, this simplifies to $\mu(t) = t$.
532:
533: Equation (\ref{v2}) gives the membrane dynamics of the current spiking
534: cycle in response to synaptic inputs from all the previous cycles.
535: We can evaluate the
536: combined synaptic input from all of the previous spikes if we assume
537: periodic firing. For the
538: neuron spiking in the past
539: at times $t=-lT$, $l=0,1,2,\dots$, we can re-express the recursion
540: relation (\ref{recur}) as the sum
541: \begin{equation}
542: S(t)=\sum_{l=0}^\infty a^l S_f(t+lT).
543: \end{equation}
544: For the saturating synaptic model, $a=\exp(-\alpha \Delta t)$ and
545: $S_f$ is given in Eq.~(\ref{histupdate}).
546: This is a geometric series and can be summed to give
547: \begin{equation}
548: S(t)=(1-a) \frac{e^{-t/\tau}}{1- a e^{-T/\tau}}.
549: \label{syn3}
550: \end{equation}
551:
552: For model 2 (using $\mu(t) = t$) we can integrate
553: (\ref{v2}) explicitly using (\ref{syn3}) to obtain
554: \begin{equation}
555: v(T) = 1 = I(1-e^{- T})-\frac{g \tau (1-a)(e^{-T/\tau} -
556: e^{- T})}{(\tau - 1)(1-a e^{-T/\tau})}.
557: \label{transfull}
558: \end{equation}
559: Equation~(\ref{transfull}) is a transcendental relation that determines the
560: spiking period and hence frequency of the synchronous network.
561: If we can completely ignore the past (\ie~use synaptic model
562: (\ref{synupdate})) then $a \sim 0$ and we obtain
563: \begin{equation}
564: v(T) = 1 = I(1-e^{- T})-\frac{g \tau }{\tau - 1}(e^{-T/\tau} -
565: e^{- T}).
566: \label{trans}
567: \end{equation}
568: For simplicity, we will conduct our analysis on this equation.
569:
570:
571: For nonsaturating synapses, we can again calculate $S(t)$ for
572: periodic firing as we did for the previous case. Here, we find
573: \begin{equation}
574: S(t)=\frac{e^{-t/\tau}}{1-e^{-T/\tau}}.
575: \label{syn4}
576: \end{equation}
577: For model 2 (using $\mu(t) = t$) we can integrate
578: (\ref{v2}) explicitly using (\ref{syn4})
579: to obtain
580: \begin{equation}
581: v(T) = 1 = I(1-e^{- T})-\frac{g \tau (e^{-T/\tau} -
582: e^{- T})}{(\tau - 1)(1-e^{-T/\tau})}.
583: \label{transns}
584: \end{equation}
585:
586: \section{Frequency Regimes}
587:
588: In this section we describe in detail the three asymptotic
589: regimes discussed in the Introduction.
590: We derive the parameter ranges in which each regime holds,
591: and the dependence of the network period on the inhibitory
592: decay time within each regime.
593:
594: \subsection{Tonic Regime}
595: The Tonic regime occurs where
596: \begin{equation}
597: T<<1, \qquad T <<\tau.
598: \label{Tcond}
599: \end{equation}
600: Note that unity in (\ref{Tcond}) corresponds to the intrinsic
601: time scale of the membrane, which has been scaled away in
602: our normalization, and $\tau$ represents the ratio
603: of the synaptic decay time to the membrane time scale.
604: The relationship in (\ref{Tcond}) allows us to expand the
605: exponentials in Eq.~(\ref{trans}) to linear
606: order in $T$ and $T/\tau$ to obtain for saturating synapses
607: \begin{equation}
608: T \simeq (I-g)^{-1}.
609: \label{TI}
610: \end{equation}
611: To derive the condition which must be satisfied for this regime
612: we impose our original assumptions (\ref{Tcond}) on
613: Eq.~(\ref{TI}). This leads to the
614: condition
615: \begin{equation}
616: (I-g)^{-1} << \min [\tau, 1]
617: \label{Tdef}
618: \end{equation}
619: Condition
620: (\ref{Tdef}) can be considered to be the definition of the tonic
621: regime.
622: A strong applied current (large $I$) and a
623: weak synapse (small $g$) suffice to satisfy it.
624:
625:
626: In the tonic regime, the synaptic decay time does not influence the
627: period, as can be seen from Eq. (\ref{TI}).
628: The synapse only affects the
629: period through the average
630: amount of inhibition the neuron receives over the course of one
631: period. We see this immediately by replacing $S(t)$ in
632: Eq.~(\ref{v2}) by its average $\langle S(t) \rangle$.
633: After integrating, we obtain the relation
634: \begin{equation}
635: 1 \simeq [I-g \langle S(t) \rangle] (1-e^{-T}),
636: \label{avgTrans}
637: \end{equation}
638: where the time average of the synaptic input $S(t)$ is defined as
639: \begin{equation}
640: \langle S(t)\rangle = \frac{1}{T} \int_0^T S(t') dt'.
641: \label{avgS}
642: \end{equation}
643: In the tonic regime to leading order this leads to the simple result of
644: $\langle S(t) \rangle \sim 1$ for saturating synapses
645: and $\langle S(t) \rangle \sim \tau/T$ for nonsaturating synapses
646: (using Eq.~(\ref{transns})).
647: If we substitute $\langle S(t) \rangle$ into Eq.~(\ref{avgTrans})
648: and linearize in $T$ and $T/\tau$ (using (\ref{Tcond})), we obtain
649: the same period as given by Eq.~(\ref{TI}) for saturating synapses.
650: For nonsaturating synapses, we find the period obeys
651: \begin{equation}
652: T \simeq I^{-1}(1+g\tau).
653: \label{TIns}
654: \end{equation}
655: Unlike the saturating case there is linear dependence on
656: $\tau$. However, since $g/I$ is small in this regime this dependence
657: is small.
658:
659: The inclusion of the synaptic contribution to the decay rate in
660: model~1 does not alter the results.
661: In the tonic regime, $T << \tau$ so we can expand Eq.~(\ref{mu}) to
662: obtain
663: \begin{equation}
664: \mu(t) \simeq t + \gamma t\equiv r t.
665: \label{muexpand}
666: \end{equation}
667: The inclusion of the synaptic contribution to the rate
668: only changes the effective passive decay rate from 1 to $r$. Using
669: this in Eq.~(\ref{trans}) and expanding to linear order yields the the
670: same period as given in Eq.~(\ref{TI}).
671:
672:
673:
674: \subsection{Phasic Regime}
675: The phasic regime occurs where
676: \begin{equation}
677: 1<< T,\qquad 1<< \tau.
678: \label{condP}
679: \end{equation}
680: This implies that $T>>T/\tau$ which
681: allows us to ignore $e^{-T}$ with
682: respect to $e^{-T/\tau}$ and 1. In this regime the neuron fires
683: at a frequency which is slow but on
684: the order of $\tau^{-1}$. Applying condition (\ref{condP}) to
685: Eq.~(\ref{trans}) yields
686: \begin{equation}
687: 1=I - \frac{g \tau}{\tau-1} e^{-T/\tau}.
688: \end{equation}
689: for saturating synapses. Solving for $T$ gives
690: \begin{equation}
691: T = \tau \ln\left[\frac{g\tau}{(\tau-1)(I-1)}\right].
692: \label{TD}
693: \end{equation}
694: Equation~(\ref{TD}) shows that in this regime the period is
695: approximately proportional to $\tau$ and logarithmically dependent on
696: the other parameters. We called this regime the phasic regime in
697: White \et~(1997) because the synaptic variable $S(t)$ is dominated by
698: a phasic component with a period proportional to $\tau$. To derive
699: the condition on the parameters for the network to be in the phasic
700: regime, we apply the condition $T>>1$ to Eq.~(\ref{TD}). We then have
701: the condition
702: \begin{equation}
703: \ln\left[\frac{g\tau}{(\tau-1)(I-1)}\right] >> \tau^{-1}.
704: \label{Pdef}
705: \end{equation}
706: Thus, the phasic regime is obtained for $g$ large compared to $I-1$.
707: We should note that $I>1$ is the condition required for firing and
708: thus must always be satisfied.
709:
710: We can compare
711: condition (\ref{Pdef}) to condition (\ref{Tdef}) for the tonic
712: regime.
713: Both the phasic and the tonic regimes require fairly large
714: $\tau$ but it is the relationship between $g$ and $I$ that
715: distinguishes the two regimes. The phasic regime is attained by a strong
716: synapse
717: and a weak applied current and the tonic regime is attained by the
718: opposite.
719:
720: For nonsaturating synapses using Eq.~(\ref{transns}) the period takes
721: the form
722: \begin{equation}
723: T = \tau \ln\left[\frac{g\tau + (\tau-1)(I-1)}
724: {(\tau-1)(I-1)}\right].
725: \label{TDns}
726: \end{equation}
727: The period is again proportional to $\tau$ but the logarithmic factor
728: is changed.
729:
730: The period is modified for model~1 but remains
731: proportional to $\tau$. From Eq.~(\ref{mu}) we can see that
732: $\mu(t)$ is bounded by $r t > \mu(t) > t$, where $r=1 + \gamma > 1$.
733: Hence the period is bounded between (\ref{TD}) and
734: \begin{equation}
735: T=\tau\ln\left[\frac{r g \tau}{(r\tau-1)(I-r)}\right].
736: \end{equation}
737: Thus, $T\propto \tau$ but the proportionality
738: constant is changed by a logarithmic factor. Here $I>r$ must be
739: satisfied
740: for the neuron to fire.
741:
742:
743: \subsection{Fast Regime}
744: The phasic and tonic regimes are applicable for slow synapses, where
745: the synaptic decay time is comparable to or slower than the period.
746: Here, we investigate the fast regime where the synapse is very fast
747: compared to the period; \ie
748: \begin{equation}
749: \tau<< T.
750: \label{condF}
751: \end{equation}
752: The fast regime can occur for a range of $\tau$.
753:
754: We first consider $\tau<<1$ so that
755: we can ignore $e^{-T/\tau}$ with
756: respect to $e^{-T}$.
757: In this case Eq.~(\ref{trans}) for saturating synapses becomes
758: \begin{equation}
759: 1\simeq I(1-e^{- T}) - g\tau e^{- T},
760: \end{equation}
761: where we have also expanded to linear order in $\tau$. From this
762: we obtain
763: \begin{equation}
764: T= \ln\left[\frac{g\tau+I}{I-1}\right].
765: \label{T3}
766: \end{equation}
767: Note that the network period is a logarithmic
768: function of the parameters.
769: Thus, in this regime, the period is dominated by the membrane
770: time constant, which is implicit in our scaling.
771: For this regime to be valid we must impose condition (\ref{condF}) on
772: (\ref{T3}). Thus we have the condition
773: \begin{equation}
774: \ln\left[\frac{g\tau+I}{I-1}\right]>>\tau.
775: \label{Fdef}
776: \end{equation}
777: This can be satisfied for $I\sim 1$ for arbitrary $g$ or for $g$ large
778: enough and arbitrary $I>1$.
779: Note that if $\tau$ is decreased from either the tonic or phasic
780: regimes (see Eq.~(\ref{Tdef}) and Eq.~(\ref{Pdef})), Eq.~(\ref{Fdef})
781: is still satisfied, implying that the network can enter the fast
782: regime from either the phasic or tonic regimes. This occurs because
783: the relationship between $g$ and $I$ distinguishes the latter two
784: regimes, while the condition for the fast regime is primarily
785: dependent on $\tau$.
786:
787: For $\tau$ near to 1, Eq.~(\ref{T3}) is no longer valid since we
788: cannot ignore $e^{-T/\tau}$ with respect to $e^{-T}$. By going back
789: to (\ref{trans}) for $\tau\sim 1$ and $T>>1$
790: we get the new period relation
791: \begin{equation}
792: 1 \simeq I(1-e^{-T}) - g T e^{-T}
793: \end{equation}
794: which again can be satisfied (since $T>>\tau\sim 1$) for $I\sim 1$ and
795: arbitrary $g$ or for $g$ large
796: enough and arbitrary $I>1$.
797:
798: The period is identical for nonsaturating synapses since the synapse
799: is so fast it does not have a chance to reach saturation.
800: The period also does not
801: change much for model 1. This is because the synapse
802: decays quickly (\ie~$\tau<<T$). Thus over most of the period,
803: $\mu(t)\sim t$ is a good approximation to Eq. (\ref{mu}).
804:
805: \section{Relationship between Full and Reduced Models: Two
806: Examples}\label{example}
807:
808: The three regimes analyzed above have been observed in numerical
809: simulations of conductance-based models. The phasic and tonic regimes
810: were reported in \citeauthor{white97}~(\citeyear{white97}) and the
811: fast regime, for spiking and/or bursting neurons, has been observed
812: previously~\cite{friesen94,perkel74,skinner94,vrees94,wang92}. Here,
813: we show that a stronger correspondence can be drawn between the
814: reduced and the conductance-based models. We show that the period for
815: the reduced model as a function of the parameters
816: $\tilde{\tau}$, $\tilde{I}$, and $\tilde{g}$ approximates the period
817: obtained from simulations of the two conductance-based neurons given
818: in the Appendix. Specifically we show that the period data
819: $\tilde{T}=\tilde{T}(\tilde{I},\tilde{g},\tilde{\tau})$ obtained from
820: the simulations of the conductance based models is well approximated
821: by the period function for the reduced models obtained from the
822: relation (\ref{transfull}).
823:
824: The period relation (\ref{transfull}) for the reduced model is in
825: terms of dimensionless quantities $T$, $I$, $g$, and $\tau$
826: (\ie~$T=T(I,g,\tau)$) . To compare to the numerical results we
827: restore the
828: dimensions using the scaling transformations given in
829: (\ref{transform}), namely
830: \begin{equation}
831: \tilde{T}=\frac{T}{\tau_m}, \qquad I=\frac{\tilde{I}+ I_r}{I_T},
832: \qquad g= \frac{\tilde{g}}{g_T}.
833: \label{invtrans}
834: \end{equation}
835: Our task therefore is to find a single set of scaling parameters
836: $I_r$, $I_T$, $g_T$ and $\tau_m$ for which the reduced model period
837: matches that of the conductance-based model.
838:
839: We first numerically generated (for each of the two models given in
840: the appendix) three tables of period data. In the first table, we
841: listed $\tilde{T}$ as $\tilde{I}$ varied for several fixed values of
842: $\tilde{g}$ and $\tilde{\tau}$, while the other two tables similarly
843: listed $\tilde{T}$ as $\tilde{g}$ or $\tilde{\tau}$ varied
844: respectively. Each table was then considered a numerical function,
845: $\tilde{T}_{(\tilde{g},\tilde{\tau})}(\tilde{I})$,
846: $\tilde{T}_{(\tilde{I},\tilde{\tau})}(\tilde{g})$ or
847: $\tilde{T}_{(\tilde{I},\tilde{g})}(\tilde{\tau})$, and we looked
848: for a choice of scaling parameters (\ref{invtrans}) such that the
849: reduced model period function $T=T(I,g,\tau)$ implicit in
850: (\ref{transfull}) was closest to the
851: numerical functions. We defined a measure of error as the sum of the
852: absolute values of the maximal deviation between the reduced and
853: conductance-based models (\ie~${\rm err}=\sum \max | \tilde{T} - T
854: |$, where the maximum was taken over all period values within a table
855: and the sum was taken over the three tables). The fit error between
856: the conductance-based and the reduced model was then minimized using
857: the function FMINS in the software package MATLAB. Critical to the
858: success of the fitting procedure was finding a good initial guess.
859: This was obtained by adjusting the scaling parameters by hand and
860: fitting the functions by eye.
861:
862: The memory function $a=\exp(\tilde{\alpha}\Delta\tilde{t})$ was fixed
863: for each conductance-based model. $\Delta \tilde t$ was defined as the spike
864: width at $V=0$, while both models used $\tilde{\alpha}=1$. The
865: results of the fits for both models are shown in Fig.~\ref{fig:freq1}.
866: Note that each neuron model has a single transformation set which is
867: used in all three panels. The phasic and tonic regimes are evident in
868: the figures. We did not explore the fast regime. The period derived
869: from the reduced model (\ref{transfull}) is seen to capture quite well
870: the dependence of the period upon the parameters of the
871: conductance-based models. Note that the White~\et~model has a
872: different spike shape from the reduced Traub and Miles model (see
873: Fig.~\ref{fig:spike}). The latter has a significant negative
874: overshoot while the former does not. Nevertheless, the quality of the
875: fit is equally good in both cases.
876:
877: The fit tends to work best in the phasic regime. This is expected
878: because the period behavior is dominated by the choice of synaptic
879: model (which is the same in the analysis and simulations). In the
880: tonic regime, the effect of the synapse is weak and the frequency
881: characteristics are more like that of the uncoupled neuron. Thus
882: differences in the intrinsic membrane dynamics will be the most
883: prominent here. The period of the uncoupled reduced model obeys the
884: classic result $T= \ln(I/I-I^*)$~\cite{tuckwell}, and we showed in the
885: tonic regime that the synapse affects the period only by changing the
886: amount of tonic input the neuron receives. The period of the Type I
887: neurons in the sense of \citeauthor{bard96}~(\citeyear{bard96}) behave
888: as $T=(I-I^*)^{-1/2}$ near the bifurcation point. Thus at least near
889: the onset of firing, the frequency behavior as a function of $I$
890: should not be the same.
891:
892: As the synapse becomes stronger, the influence of the synaptic
893: dynamics begins to compete with the intrinsic membrane dynamics. Deep
894: within the phasic regime the synaptic dynamics completely dominate and
895: the period is primarily determined by the response of the neuron to
896: inhibition. Our analysis showed that the intrinsic dynamics are not
897: important in the phasic regime. However, the use of model 2 instead
898: of model 1 for equation (\ref{transfull}) will result in quantitative
899: inaccuracies even well within in the phasic regime.
900:
901:
902: \section{Discussion and Conclusions}
903:
904: In this paper, we show that a fully synchronized network can have its
905: frequency determined by several parameters, including applied
906: currents, synaptic strength and synaptic decay time. However, in
907: regimes other than the phasic regime, in which parameters other than
908: the synaptic decay dominate the network frequency, even mild
909: heterogeneity can eliminate coherence~\cite{white97}. Indeed, in that
910: work we showed that increasing the drive $I$, which increases the
911: uncoupled frequency of each cell or the common frequency of a
912: homogenous inhibitory network, can desynchronize a mildly
913: heterogeneous network. This is true even if the percentage difference
914: in natural frequencies over the population is held constant.
915: Decreasing the drive sufficiently puts the network in a regime where
916: inhibition suppresses spiking in less excitable cells. This also
917: causes a loss of coherence albeit through a different mechanism.
918: These two effects limit the possible frequencies at which the network
919: can be synchronized, and the time scale of the synapse plays a crucial
920: role in determining the width of this frequency band.
921:
922: With respect to the hippocampus, the analysis
923: predicts the fast 200 Hz
924: ripples observed in CA1~\cite{ylinen95}, cannot be solely mediated by
925: GABA$_A$ synaptic inhibition, whose time scales are much slower than
926: 5~ms. An additional mechanism such as
927: gap junction mediated electrical coupling may be responsible for
928: the observed synchronous fast ripples.
929: The 40 Hz gamma rhythm, which is on the order of the time
930: scale of GABA$_A$, can be supported by inhibition alone. This was
931: observed in experiments in hippocampal
932: slices~\cite{whit95,traub96,jeff96}. Our results on
933: the dependence of the period on the synaptic decay time constant
934: clarifies the experimental
935: observations. Our analysis predicts that for
936: coherent oscillations, the network should be in the phasic regime.
937: To compare with the analysis, we replotted the experimental data of
938: \citeauthor{whit95}~(\citeyear{whit95}) as {\em
939: period} versus synaptic decay time in Fig.~\ref{fig:jeff}. We note
940: that the period
941: $T$ is proportional to the decay time $\tau$, as predicted for the
942: phasic regime. Furthermore, $\tau / T$ has a value less than 1,
943: consistent with the requirement for being in the phasic regime.
944:
945:
946: For homogeneous networks, the time scale of
947: the inhibition is not critical for the frequency of a synchronized
948: network. A homogeneous network can stably synchronize at parameter
949: regimes outside the phasic regime. However, for a mildly
950: heterogeneous network, the synchronization mechanism must not only
951: draw together phases, but must also help create a common frequency.
952: As shown in~\cite{chow97}, inhibitory synapses in the phasic regime
953: {\em do} provide such a mechanism for heterogeneous networks.
954: We wish to point out that this frequency dependent synchronization is
955: not necessary for all networks. For instance,
956: consider a network coupled with (nonrectified) electrical coupling.
957: In the fully synchronized state,
958: the coupling currents disappear and hence do not contribute to the
959: network frequency. With mild heterogeneity, the network still can maintain
960: coherence, but there is no time constant to
961: affect the network frequency.
962:
963:
964: The connection of the frequency of a synchronized network with
965: inhibition to the inhibitory decay time has previously been understood
966: intuitively~\cite{destexhe93,skinner93,kopell,wang96}. Here we
967: present a simple example to clarify how all of the relevant time
968: scales interact to produce the regimes in which that intuition is
969: correct. These conclusions assume that the membrane potential between
970: spikes is governed by a passive decay process that approaches a fixed
971: threshold. The reduction from model 1 to model 2 simplifies the
972: effect of the synapse. The accuracy of our analytical
973: calculations depend on the validity of these approximations.
974: Nevertheless, we expect the three regimes to still exist for a much
975: larger class of conductance based models and models of synapses.
976: Evidence for this conjecture was given in Section~\ref{example}.
977: We note that~\citeauthor{IH97}~(\citeyear{IH97}) have shown that for weakly
978: coupled systems an arbitrary Type I excitable neuron can be
979: transformed into integrate-and-fire form by a piece-wise continuous change
980: of variables. Our results show that this correspondence can hold
981: beyond the weak coupling limit.
982:
983:
984: \acknowledgements
985: We wish to thank B. Ermentrout for many clarifying discussions and
986: help in computing phase response curves.
987: We also thank S. Epstein, W. Gerstner, C. Linster,
988: J. Rinzel, C. Soto-Trevi\~no, and R. Traub, for helpful discussions.
989: We thank J.~Jefferys for providing us with experimental data.
990: This work was supported in part by
991: the National Science Foundation (DMS-9631755 to N.K. and J.W.),
992: the National Institutes of Health (MH47150 to N.K.;
993: 1R29NS34425 to J.W.),
994: and The Whitaker Foundation (to J.W.)
995:
996:
997:
998: \appendix
999:
1000: \section{Neuron Dynamics}\label{app:neu}
1001:
1002: For our physiologically based neuron we consider a
1003: single-compartment model
1004: with inhibitory synapses obeying first-order kinetics.
1005: The membrane potential obeys the current balance
1006: equation
1007: \begin{equation}
1008: C\frac{dV_i}{dt}=\tilde{I}-I_{Na}-I_K-I_L-I_s,
1009: \label{membrane}
1010: \end{equation}
1011: where $C=1 \mu$F/cm$^2$, $\tilde{I}$ is the applied current,
1012: $I_{Na}=g_{Na} m^3 h (V_i - V_{Na})$ and $I_K=g_Kn^4(V_i-V_K)$ are
1013: the spike generating currents, $I_L= g_L(V_i-V_L)$
1014: is the leak current and $I_s = \sum_j^N \tilde{g} S_j(t) (V_i-V_s)$ is the
1015: synaptic current.
1016:
1017: The interneuron model in White \et~(\citeyear{white97}) used parameters:
1018: $g_{Na}=30$~mS/cm$^2$, $g_K=20$~mS/cm$^2$,
1019: $g_L=0.1$~mS/cm$^2$,
1020: $V_{Na}=45$~mV, $V_K=-80$~mV, $V_L=-60$~mV, $V_s=-75$~mV.
1021: The activation variable $m$ was assumed fast and substituted with
1022: its asymptotic value
1023: $m=m_\infty(v)=(1 + \exp[-0.08(v+26)])^{-1}$. The gating variables $h$
1024: and $n$ obey
1025: \begin{equation}
1026: \frac{dh}{dt}=\frac{h_\infty(v)-h}{\tau_h(v)},
1027: \qquad
1028: \frac{dn}{dt}=\frac{n_\infty(v)-n}{\tau_n(v)},
1029: \end{equation}
1030: with
1031: $h_\infty(v)=(1 + \exp[0.13(v+38)])^{-1}$,
1032: $\tau_h(v)=0.6/(1 + \exp[-0.12(v+67)])$,
1033: $n_\infty(v)=(1 + \exp[-0.045(v+10)])^{-1}$, and
1034: $\tau_n(v)=0.5+2.0/(1+\exp[0.045(v-50)])$.
1035:
1036: The reduced Traub and Miles
1037: model~\cite{ermenkopell98,traub96b} used
1038: parameters: $g_{Na}=100$~mS/cm$^2$,
1039: $g_K=80$~mS/cm$^2$, $g_L=0.1$~mS/cm$^2$,
1040: $V_{Na}=50$~mV, $V_K=-100$~mV, $V_L=-67$~mV, $V_s=-80$~mV;
1041: $m=m_\infty(v)=\tilde{\alpha}_m(v)/(\tilde{\alpha}_m(v)+\tilde{\beta}_m(v))$,
1042: where
1043: $\tilde{\alpha}_m(v)=0.32 (54+v)/(1-\exp(-(v+54)/4))$ and
1044: $\tilde{\beta}_m(v)=0.28 (v+27)/(\exp((v+27)/5)-1)$;
1045: \begin{equation}
1046: \frac{dn}{dt}=\tilde{\alpha}_n(v)(1-n)-\tilde{\beta}_n(v)n
1047: \end{equation}
1048: with $\tilde{\alpha}_n(v)=0.032(v+52)/(1-\exp(-(v+52)/5))$,
1049: $\tilde{\beta}_n(v)=0.5\exp(-(v+57)/40)$;
1050: $h=h_\infty(v)=\max[1 - 1.25 n, 0]$.
1051:
1052:
1053: For both models,
1054: the synaptic gating variable $S_j(t)$ is assumed to obey first order
1055: kinetics of the form
1056: \begin{equation}
1057: \frac{dS_j}{dt} = \tilde{{\alpha}} F(V_j)(1-S_j)-\tilde{{\beta}} S_j,
1058: \end{equation}
1059: where $\tilde{{\alpha}}=1$ and $\tilde{{\beta}}$ are
1060: respectively the
1061: synaptic rise and decay rates and $F(V_i)=1/(1+\exp[-V_i])$.
1062: Transmission delays are neglected. We are interested in the response
1063: of the system to changes in the applied current $\tilde{I}$, synaptic
1064: conductance $\tilde{g}$, and synaptic decay time $\tilde{\tau}\equiv
1065: \tilde{{\beta}}^{-1}$. The ODEs were integrated using either a fourth-order
1066: Runge-Kutta method or the Gear method.
1067:
1068:
1069: \begin{thebibliography}{}
1070:
1071: \bibitem[\protect\citeauthoryear{Abbott and van
1072: Vreeswijk}{1993}]{abbott93} Abbott L, and van
1073: Vreeswijk C. Asynchronous states
1074: in networks of pulse-coupled oscillators. {\it Phys. Rev. E} 48:1483-1490.
1075:
1076: \bibitem[\protect\citeauthoryear{Bragin \et }{1995}]{bragin95}
1077: Bragin A, Jand\'o G, N\'adasdy Z, Hetke J, Wise K,
1078: and Buzs\'aki G. (1995) Gamma (40-100 Hz) oscillation in the hippocampus of
1079: the behaving rat. {\it J. Neuroscience} 15:47-60.
1080:
1081: \bibitem[\protect\citeauthoryear{Buzs\'aki}{1986}]{buzsaki86}
1082: Buzs\'aki G. (1986) Hippocampal sharp waves: their origin
1083: and significance. {\it Brain Res.} 398:242-252.
1084:
1085: \bibitem[\protect\citeauthoryear{Chow}{1998}]{chow97} Chow
1086: CC. Phase-locking in weakly heterogeneous neuronal networks. Physica D
1087: (in press).
1088:
1089: \bibitem[\protect\citeauthoryear{Destexhe \et}{1993}]{destexhe93}
1090: Destexhe A, McCormick DA and Sejnowski TJ. (1993) A model for 8-10 Hz
1091: spindling in interconnected thalamic relay and reticularis
1092: neurons. {\it Biophysical J.} 65:2473-2477.
1093:
1094: \bibitem[\protect\citeauthoryear{Destexhe \et}{1994}]{destexhe94}
1095: Destexhe A, Mainen ZF and Sejnowski TJ. (1994) An efficient method for
1096: computing synaptic conductances based on a kinetic model of receptor
1097: binding. {\it Neural Comp.} 6:14-18.
1098:
1099: \bibitem[\protect\citeauthoryear{Ermentrout}{1996}]{bard96}
1100: Ermentrout GB. (1996) Type I membranes, phase resetting curves, and
1101: synchrony. {\it Neural Comp.} 8:979-1001.
1102:
1103: \bibitem[\protect\citeauthoryear{Ermentrout and
1104: Kopell}{1998}]{ermenkopell98} Ermentrout GB and Kopell N. (1998)
1105: Fine structure of neural spiking and synchronization in the presence
1106: of conduction delays. {\it Proc.~Nat.~Acad.~Sci.~USA} 95:1259-1264.
1107:
1108: \bibitem[\protect\citeauthoryear{Friesen}{1994}]{friesen94}
1109: Friesen W. (1994) Reciprocal inhibition, a mechanism underlying
1110: oscillatory animal movements. {\it Neurosci Behavior} 18:547-553.
1111:
1112: \bibitem[\protect\citeauthoryear{Gerstner \et}{1996}]{gerstner96}
1113: Gerstner W, van Hemmen JL, and Cowen J. (1996) What
1114: matters in neuronal locking? {\it Neural Comp.} 8:1653-1676.
1115:
1116: \bibitem[\protect\citeauthoryear{Golomb and Rinzel}{1993}]{golomb93}
1117: Golomb D and Rinzel J. (1993) Dynamics of globally coupled inhibitory
1118: neurons with heterogeneity. Phys. Rev. E {\bf 48}, 4810-4814.
1119:
1120: \bibitem[\protect\citeauthoryear{Gray}{1994}]{gray94} Gray CM. (1994)
1121: Synchronous oscillations in neuronal
1122: systems: mechanisms and functions. {\it J. Comp. Neuro.} 1:11-38.
1123:
1124: \bibitem[\protect\citeauthoryear{Hansel \et}{1995}]{hansel95} Hansel
1125: D, Mato G, and Meunier C. (1995) Synchrony in
1126: excitatory neural networks. {\it Neural Comp.} 7:307-337.
1127:
1128: \bibitem[\protect\citeauthoryear{Izhikevich and
1129: Hoppensteadt}{1997}]{IH97} Izhikevich EM. and Hoppensteadt F. {\em Weakly
1130: Connected Neural Networks}, (Springer-Verlag, 1997, Chapter 8).
1131:
1132: \bibitem[\protect\citeauthoryear{Jefferys \et}{1996}]{jeff96} Jefferys
1133: JGR, Traub RD, and Whittington MA. (1996)
1134: Neuronal networks for induced `40 Hz' rhythms. {\it Trends Neurosci.}
1135: 19:202-207.
1136:
1137: \bibitem[\protect\citeauthoryear{Katsumaru \et}{1988}]{katsu}
1138: Katsumaru H, Kosaka T, Heizman CW, and Hama K. (1988)
1139: Gap-junctions on GABAergic neurons containing the calcium-binding
1140: protein parvalbumin in the rat hippocampus (CA1 regions).
1141: {\it Exp. Brain Res.} 72:363-370.
1142:
1143: \bibitem[\protect\citeauthoryear{Kopell and LeMasson}{1994}]{kopell}
1144: Kopell N. and LeMasson G. (1994) Rhythmogenesis, amplitude modulation,
1145: and multiplexing in a cortical architecture. {\it
1146: Proc. Natl. Acad. Sci. USA} 91:10586-10590.
1147:
1148: \bibitem[\protect\citeauthoryear{Llin\'as and Ribary}{1993}]{llinas93}
1149: Llin\'as R and Ribary U. (1993) Coherent 40-Hz oscillation
1150: characterizes dream state in humans. {\it Proc. Natl. Acad. Sci. USA}
1151: 90:2078-2081.
1152:
1153: \bibitem[\protect\citeauthoryear{Perkel and Mulloney}{1974}]{perkel74}
1154: Perkel D and Mulloney B. (1974) Motor patterns in reciprocally
1155: inhibitory neurons exhibiting postinhibitory rebound.
1156: {\it Science} 185:181-183.
1157:
1158: \bibitem[\protect\citeauthoryear{Pinsky and Rinzel}{1994}]{pinsky}
1159: Pinsky P and Rinzel J. (1994) Intrinsic and network
1160: rhythmogenesis in a reduced Traub model for CA3
1161: neurons. {\it J. Comp. Neurosci.} 1:39-60.
1162:
1163: \bibitem[\protect\citeauthoryear{Rinzel and Ermentrout}{1989}]{rinzel}
1164: Rinzel and Ermentrout, in {\em Methods in Neuronal
1165: Modelling}, edited by C. Koch and I. Segev (MIT Press, Cambridge,
1166: 1989).
1167:
1168: \bibitem[\protect\citeauthoryear{Ritz and Sejnowski}{1997}]{ritz97}
1169: Ritz R and Sejnowski TJ. (1997) Synchronous oscillatory activity in
1170: sensory systems: new vistas on mechanisms. {\it Curr. Opin. Neurobio.}
1171: 7:536-546.
1172:
1173: \bibitem[\protect\citeauthoryear{Skinner \et}{1993}]{skinner93}
1174: Skinner F, Turrigiano GG, and Marder E. (1993) Frequency and burst
1175: duration in oscillating neurons and two-cell networks. {\it Biol. Cybern.}
1176: 69:375-383.
1177:
1178: \bibitem[\protect\citeauthoryear{Skinner \et}{1994}]{skinner94}
1179: Skinner F, Kopell N, and Marder E. (1994) Mechanisms for oscillations
1180: and frequency control in networks of mutually inhibitory relaxation
1181: oscillators. {\it J. Comp. Neurosci.} 1:69-87.
1182:
1183: \bibitem[\protect\citeauthoryear{Terman \et}{1996}]{terman96} Terman
1184: D, Bose A, and Kopell N. (1996)
1185: Functional reorganization in thalamocortical networks: transition between
1186: spindling and delta sleep rhythms. {\it Proc. Natl. Acad. Sci. USA}
1187: 93:15417-15422.
1188:
1189: \bibitem[\protect\citeauthoryear{Terman \et}{1997}]{terman} Terman,
1190: D., Kopell N. and Bose A. Dynamics of two mutually coupled slow
1191: inhibitory neurons. {\it Physica D} (in press).
1192:
1193: \bibitem[\protect\citeauthoryear{Tuckwell}{1989}]{tuckwell} Tuckwell
1194: HC (1989) {\it Stochastic processes in the neurosciences} (SIAM,
1195: Philadelphia).
1196:
1197: \bibitem[\protect\citeauthoryear{Traub}{1995}]{traub95} Traub
1198: RD. (1995) Model of synchronized population bursts
1199: in electrically coupled interneurons containing active dendritic
1200: conductances. {\it J. Comp. Neurosci.} 2:283-289.
1201:
1202: \bibitem[\protect\citeauthoryear{Traub \et}{1996a}]{traub96} Traub RD,
1203: Whittington M, Colling S, Buzs\'aki G,
1204: and Jefferys JGR. (1996a)
1205: Analysis of gamma rhythms in the rat hippocampus in vitro
1206: and in vivo. {\it J. Physiol.} 493:471-484.
1207:
1208: \bibitem[\protect\citeauthoryear{Traub \et}{1996b}]{traub96b} Traub
1209: RD, Jefferys JGR, and Whittington MA. (1996b)
1210: Simulation of gamma rhythms in networks of
1211: interneurons and pyramidal cells. {\it J. Comp. Neurosci.} In press.
1212:
1213: \bibitem[\protect\citeauthoryear{Tsodyks \et}{1993}]{tsodyks93}
1214: Tsodyks M, Mitkov I, and Sompolinsky, H. (1993) Pattern of synchrony
1215: in inhomogeneous networks of oscillators with pulse interactions.
1216: {\it Phys. Rev. Lett.} 71:1280-1283.
1217:
1218: \bibitem[\protect\citeauthoryear{van Vreeswijk \et}{1994}]{vrees94}
1219: van Vreeswijk C, Abbott L, and
1220: Ermentrout GB. (1994) When inhibition not excitation synchronizes neural
1221: firing. {\it J. Comp. Neuroscience} 1:313-321.
1222:
1223: \bibitem[\protect\citeauthoryear{Wang and Buzs\'aki}{1996}]{wang96}
1224: Wang X-J and Buzs\'aki G. (1996) Gamma oscillation by
1225: synaptic inhibition in an interneuronal network model.
1226: {\it J. Neurosci.} 16:6402-6413.
1227:
1228: \bibitem[\protect\citeauthoryear{Wang and Rinzel}{1992}]{wang92}
1229: Wang X-J and Rinzel J. (1992) Alternating and synchronous rhythms in
1230: reciprocally inhibitory model neurons. {\it Neural Comp.} 4:84-97.
1231:
1232: \bibitem[\protect\citeauthoryear{Whittington \et}{1995}]{whit95}
1233: Whittington MA, Traub RD, and Jefferys JGR. (1995)
1234: Synchronized oscillations in interneuron networks
1235: driven by metabotropic glutamate receptor activation.
1236: {\it Nature} 373:612-615.
1237:
1238: \bibitem[\protect\citeauthoryear{White \et}{1998}]{white97} White JA,
1239: Chow CC, Ritt J, Soto-Trevi\~no C, and Kopell N. (1998)
1240: Synchronization and oscillatory dynamics in heterogeneous, mutually
1241: inhibited neurons. {\it J. Comp. Neurosci.} (in press).
1242:
1243: \bibitem[\protect\citeauthoryear{Ylinen \et}{1995}]{ylinen95} Ylinen
1244: A, Bragin A, N\'adasdy Z, Jand\'o G, Szab\'o I,
1245: Sik A, and Buzs\'aki G. (1995) Sharp wave-associated high-frequency
1246: oscillation (200 Hz) in the intact hippocampus: network and
1247: intracellular mechanisms. {\it J. Neurosci.} 15:30-46.
1248:
1249: \end{thebibliography}
1250:
1251: \begin{figure}
1252: \centerline{\epsfig{figure=fig1.eps, height=7.in,bbllx=14pt,bblly=9pt,bburx=597pt,bbury=784pt, clip=}}
1253: \caption{Example voltage traces of spikes for the A)
1254: White~\et~(1998) interneuron and B) reduced
1255: Traub and Miles model of Ermentrout and Kopell~(1998).}
1256: \label{fig:spike}
1257: \end{figure}
1258:
1259: \begin{figure}
1260: \centerline{\epsfig{figure=fig2.eps, height=7.in,bbllx=14pt,bblly=9pt,bburx=597pt,bbury=784pt, clip=}}
1261: \caption{Shown are example slices of the numerically determined period
1262: for the conductance-based models (dashed lines) compared to the
1263: period of the reduced model
1264: obtained from Eq.~(\ref{transfull}) (solid lines). Figures A), B)
1265: and C) show results for the White \et (1998) model with memory
1266: coefficient $a=0.30$, and scaling parameters $I_r=1.9155$, $I_T=1.4337$,
1267: $\tau_m=12.0230$, and $g_T=0.0851$. Figures D), E) and F) show
1268: results for the
1269: reduced Traub and Miles neuron with $a=0.74$, $I_r=1.3546$, $I_T=1.6211$,
1270: $\tau_m=16.1158$, and $g_T=0.1111$. From top to bottom the slices
1271: have parameters:
1272: A) $(\tilde{g},\tilde{\tau})=(2,50), (1,32.5), (1,15), (0.05,5)$
1273: B) $(\tilde{I},\tilde{g})=(1.64,1),(5,1),(5,0.5),(5,0.05)$,
1274: C) $(\tilde{I},\tilde{\tau})=(1.74,50),(1.74,32.5),(1.74,15),(5,15)$,
1275: D) $(\tilde{g},\tilde{\tau})=(2,50),(1,32),(1,15),(0.5,5)$
1276: E) $(\tilde{I},\tilde{g})=(1.085,1),(4.825,1.25),(4.825,0.5),(4.825,0.05)$
1277: and F) $(\tilde{I},\tilde{\tau})=(2.02,50),(2.02,32.5),(2.02,15),(4.825,15)$.
1278: The phasic and tonic regimes are most clearly distinguished in B) and E)
1279: where the period is proportional to $\tilde{\tau}$ in the upper curves
1280: and independent of $\tau$ in the lower curves.}
1281: \label{fig:freq1}
1282: \end{figure}
1283:
1284:
1285: \begin{figure}
1286: \centerline{\epsfig{figure=fig3.eps,height=8.in,bbllx=14pt,bblly=9pt,bburx=597pt,bbury=784pt,clip=}}
1287: \caption{Plot of period vs. synaptic decay time for a hippocampal
1288: slice showing linear behavior indicative of the phasic regime.
1289: Responses were recorded from inhibitory neurons evoked by glutamate
1290: application during a wash-in of 2 $\mu$M
1291: pentobarbital~(Whittington~\et, 1995).
1292: Experimental data is provided courtesy of J. Jefferys.}
1293: \label{fig:jeff}
1294: \end{figure}
1295:
1296:
1297: \end{document}
1298:
1299: