1: % Note:

2: % 1. draft option in documentclass suppresses figures from appearing

3: % 2. Using hyperref to allow active links to references turns

4: %    off the collapsing of successive references in a

5: %    single citation. So use hyperref only for

6: %    preprints that will be viewed onscreen, not

7: %    printed out.

8: 

9: % \documentclass[preprint,pre,showpacs,superscriptaddress]{revtex4}

10: %\documentclass[pre,showpacs,superscriptaddress]{revtex4}

11: \documentclass[twocolumn,showpacs,superscriptaddress]{revtex4}

12: 

13: \usepackage{amsmath}

14: \usepackage{amssymb}

15: \usepackage{graphicx}

16: %\usepackage[colorlinks=true,linkcolor=blue]{hyperref}

17: 

18: \newcommand{\hvcra}{HVC${}_\textrm{RA}$}

19: \newcommand{\cai}{\ensuremath{\left[{\rm Ca}\right]_i}}

20: 

21: \begin{document}

22: 

23: 

24: \title{Stable Propagation of a Burst Through a

25:  One-Dimensional Homogeneous Excitatory Chain

26:  Model of Songbird Nucleus HVC}

27: 

28: \author{MengRu Li}

29: \affiliation{Department of Physics, Duke University, Durham

30: NC 27708-0305}

31: 

32: \author{Henry Greenside}

33: \affiliation{Department of Physics, Duke University, Durham

34: NC 27708-0305}

35: 

36: \date{\today}

37: 

38: \begin{abstract}

39:   We demonstrate numerically that a brief burst

40:   consisting of two to six spikes can propagate in a

41:   stable manner through a one-dimensional homogeneous

42:   feedforward chain of non-bursting neurons with

43:   excitatory synaptic connections. Our results are

44:   obtained for two kinds of neuronal models, leaky

45:   integrate-and-fire (LIF) neurons and Hodgkin-Huxley

46:   (HH) neurons with five conductances.  Over a range of

47:   parameters such as the maximum synaptic conductance,

48:   both kinds of chains are found to have multiple

49:   attractors of propagating bursts, with each attractor

50:   being distinguished by the number of spikes and total

51:   duration of the propagating burst.  These results

52:   make plausible the hypothesis that sparse

53:   precisely-timed sequential bursts observed in

54:   projection neurons of nucleus~HVC of a singing zebra

55:   finch are intrinsic and causally related.

56: \end{abstract}

57: 

58: \maketitle

59: 

60: 

61: \section{Introduction}

62: \label{sec:intro}

63: 

64: A question of great interest to neurobiology is how

65: animals learn to generate temporal patterns of muscle

66: activation. An example that has been much studied in

67: recent years because of its relevance to human

68: speech~\cite{Doupe99} and because of the rich variety

69: of possible experiments is how songbirds learn to sing

70: a song by auditory-guided vocal

71: feedback~\cite{Catchpole95}. A young male bird first

72: memorizes the song of an adult male of the same

73: species. Then over many months, over many iterations

74: (more than 50,000 iterations for a zebra

75: finch~\cite{Johnson02}), and by just listening to his

76: own vocalizations, a young male learns to vary the

77: activation of its respiratory and syringeal muscles

78: until his song is able to match accurately the original

79: memorized song. This is an impressive feat given that

80: an adult song might last several seconds, that there is

81: auditory structure that lasts less than 10~ms, that

82: many syringeal and respiratory muscles need to be

83: coordinated, and that some species of songbirds are

84: able to learn and sing many different songs. Nearly all

85: details of this process remain poorly understood, in

86: particular how the young male memorizes a complex song,

87: and how auditory feedback is used to adjust the pattern

88: of muscle activation until the songbird is able to

89: reproduce accurately the memorized song.

90: 

91: Studies of songbirds have shown that certain

92: anatomically and physiologically distinct brain regions

93: called nuclei are associated with the recognition,

94: learning, and production of song (see

95: Fig.~\ref{fig:songbird-brain}). A recent experiment by

96: Hahnloser et al~\cite{Hahnloser02} recorded

97: extracellular action potentials (spikes) from neurons

98: in awake singing male zebra finches and found that the

99: neurons in the nucleus~HVC\cite{Reiner04} that project

100: to the robust nucleus of the arcopallium (abbreviated

101: as~RA~\cite{Reiner04}) have the remarkable properties

102: of firing sparsely and precisely during singing. (In

103: the following, HVC~neurons that project to~RA will be

104: abbreviated as \hvcra~neurons.)  Typically, each

105: \hvcra~neuron fires a brief burst of about 3-4 spikes

106: once per song motif with each burst lasting about 6~ms.

107: (A motif is a cluster of distinct auditory syllables

108: that is repeated as a single pattern and that lasts an

109: average of 0.6~s for zebra finch songs.)  Measurements

110: during successive motifs from a given adult male bird

111: show that each burst from a particular \hvcra~neuron is

112: aligned with certain acoustic features of the motif to

113: a precision of about 0.7~ms~\cite{Hahnloser02}. These

114: bursts are important to understand since they are

115: believed to provide the temporal framework for

116: organizing the syllables of a song~\cite{Fee04}.

117: 

118: \begin{figure}

119:   \centering

120:   \includegraphics[width=3in]{songbird-brain.eps}

121:   \caption{ Schematic anatomical diagram of a sagittal

122:     cross section of one hemisphere of a zebra finch

123:     brain, showing the major nuclei associated with the

124:     learning and production of song.  The nucleus~HVC

125:     plays a central role since it receives auditory and

126:     other input, it drives the premotor

127:     nucleus~RA~\cite{Hahnloser02}, and it sends

128:     information to a parallel anterior forebrain

129:     pathway (nuclei~X, DLM, and~LMAN) that play a role

130:     in the learning and maintenance of a song. There

131:     are two HVC~nuclei, one on the left and one on the

132:     right side of the songbird's brain.}

133:   \label{fig:songbird-brain}

134: \end{figure}

135: 

136: The observation that each \hvcra~neuron bursts just

137: once per motif and other data which show that there is

138: a reproducible temporal ordering of \hvcra~neurons

139: according to when they fire their single burst per

140: motif~\cite{Hahnloser02}, suggest the hypothesis that

141: bursts propagate along a feedforward network of

142: excitatory \hvcra~neurons~\cite{Fee04}. The idea is

143: similar to a path of dominoes such that the falling of

144: one domino causes the next domino in the path to

145: topple. Here, a given \hvcra~neuron receives one or

146: more bursts from other \hvcra~neurons that have fired

147: recently during the motif. In response, the given

148: neuron itself fires a burst that helps to activate

149: neurons that fire later during the motif. That each

150: \hvcra~neuron fires exactly once per motif implies that

151: the connections are feedforward (although this

152: assumption can be weakened if a recurrent burst arrives

153: while some earlier neuron is still in a refractory

154: state~\cite{Jin05}). The bursts continue to propagate

155: until the chain ends and the motif stops, or perhaps

156: there is a branch point to a different chain whose

157: dynamics generate a different motif.

158: 

159: While there has been much previous theoretical research

160: concerning when spikes can propagate along excitatory

161: networks of different

162: kinds~\cite{Abeles91,Golomb97,Diesmann99,Reyes03,Litvak03,Osan04,Vogels05},

163: the songbird experiments on~HVC of Hahnloser et al

164: during singing~\cite{Hahnloser02} and related

165: experiments by Mooney of bursts in HVC projection

166: neurons during audition~\cite{Mooney00} motivate asking

167: new theoretical questions about the propagation of a

168: burst through various kinds of networks.

169: 

170: In this paper, we show that a minimal feedforward

171: architecture---a one-dimensional (1d), homogeneous,

172: excitatory chain of non-bursting neurons---can support

173: the stable propagation of a brief high-frequency burst.

174: By studying two kinds of chains, one with leaky

175: integrate-and-fire neurons (abbreviated as LIF~neurons)

176: and one with more realistic single-compartment

177: conductance-based neurons (abbreviated as HH~neurons

178: for Hodgkin-Huxley~like), we show that a burst can

179: propagate in a stable manner that does not require a

180: careful choice of neuronal model nor a careful tuning

181: of model parameters.  Depending on the values of

182: parameters such as the maximum synaptic conductance, we

183: find that each kind of 1d~chain (LIF or~HH) has

184: multiple attractors that differ in the number of spikes

185: and in the total width of the propagating burst.

186: 

187: Our calculations are the first to demonstrate that a

188: brief high-frequency burst similar to those observed

189: in~HVC can propagate in a stable manner along a simple

190: excitatory chain of neurons. (A similar result was

191: announced independently by Jin and

192: collaborators~\cite{Jin05}, who studied a more

193: complicated model. We discuss this preprint briefly in

194: Section~\ref{sec:conclusions} below.) Our results make

195: plausible one of the simplest explanations of the

196: Hahnloser et al result, namely that the observed bursts

197: of \hvcra~neurons during singing are intrinsic to~HVC

198: in that external input from other brain regions is not

199: needed to generate the bursts (except the first burst),

200: and in that the bursts are causally related such that

201: one burst initiates the generation of the next

202: burst. Our results also make some predictions such as

203: the existence of multistability (different kinds of

204: propagating bursts can occur in the same chain

205: depending on how the chain is activated), and that

206: transitions from one kind of propagating burst to

207: another kind (differing in the number of spikes and

208: burst width) can occur as parameters are varied. New

209: experimental studies of~HVC, especially using optical

210: methods with high time resolution~\cite{Ohki05} that

211: can examine the spatiotemporal activity of many neurons

212: simultaneously, may help to confirm these ideas.

213: 

214: The rest of this paper is organized as follows.

215: Section~\ref{sec:methods} discusses some details of the

216: two classes of mathematical neurons that we use in the

217: 1d~chains, and how the excitatory synaptic currents are

218: modeled.  Section~\ref{sec:results-discussion} presents

219: numerical results mainly for a 1d~chain of~HH neurons

220: (as opposed to LIF~neurons), especially the existence

221: of different attractors corresponding to different

222: kinds of propagating bursts.  The paper concludes in

223: Section~\ref{sec:conclusions}, where key results are

224: summarized. We also discuss there how our results are

225: related to alternative hypotheses such as that the

226: bursting arises from propagation through synfire

227: chains~\cite{Abeles91} or by a central pattern

228: generator (abbreviated as CPG).

229: 

230: In a second paper~\cite{Li06II,Li06thesis}, we will

231: discuss how noise and network heterogeneities affect

232: the propagation of bursts through a synfire chain. This

233: second paper shows that our key conclusions still hold:

234: a brief high-frequency burst can still propagate in a

235: stable manner for a range of noise strengths and for

236: various amounts of heterogeneity. However, we also find

237: that single spikes, even if synchronized across a

238: synfire pool, do not propagate in a stable way for most

239: of the parameters studied, which suggests that short

240: bursts of several spikes are important for achieving

241: robust propagation through realistic synfire chains.

242: 

243: 

244: \section{Methods}

245: \label{sec:methods}

246: 

247: In this section, we discuss some experimental data that

248: justifies the study of a one-dimensional chain of

249: excitatory neurons, and we also point out some of the

250: experimental details that are ignored in our model. We

251: also discuss some details of the integrate-and-fire and

252: conductance-based neuronal models with excitatory

253: synapses that we use in the

254: 1d~chains. Appendix~\ref{appendix:equations-parameters}

255: provides further details of equations and parameter

256: values.

257: 

258: \subsection{A Homogeneous 1d~Chain of Excitatory Neurons}

259: 

260: \label{sec:1d-chain-motivation}

261: 

262: The physiological properties of HVC~neurons and how

263: they are interconnected within~HVC are poorly

264: understood~\cite{Kubota98,Dutar98,Mooney00,Mooney05,Wild05}

265: so that it is not possible at this time to develop a

266: quantitative model of the HVC~microcircuitry, say at

267: the level of hippocampus

268: models~\cite{Traub99}. Researchers have

269: shown~\cite{Dutar98,Mooney00} that there are at least

270: three main classes of neurons: excitatory neurons that

271: project to nucleus~RA, excitatory neurons that project

272: to area~X in the anterior forebrain pathway, and

273: inhibitory interneurons~\cite{Dutar98,Mooney00} that

274: connect only to other neurons within~HVC. Recent

275: paired-electrode recordings by Mooney and

276: Prather~\cite{Mooney05} have shown that each type

277: of~HVC neuron makes local connections with the other

278: two types of HVC~neurons but details of the connections

279: such as the number, kinds, and strengths of synapses

280: are incompletely known.

281: 

282: Of special importance for this paper is the

283: experimental observation that \hvcra~neurons synapse

284: with other \hvcra~neurons~\cite{Mooney05}. Thus it is

285: possible for a feedforward network of excitatory

286: neurons to exist in~HVC, although we emphasize that

287: there is no evidence presently for such a

288: network. Further, there are of order~40,000

289: \hvcra~neurons in HVC~\cite{Fee04}, which should be

290: sufficiently many to create a chain whose dynamics

291: spans a motif or even several motifs. (Since the

292: observed bursts in \hvcra~neurons last

293: 6~ms~\cite{Hahnloser02}, a chain of about one hundred

294: pools of neurons would suffice to span a motif

295: of~0.6~s, provided that the bursts do not overlap in

296: time.)

297: 

298: To determine in principle whether the propagation of a

299: burst can explain the experimental

300: data~\cite{Hahnloser02,Mooney00}, we study one of the

301: simplest possible feedforward networks, namely a

302: one-dimensional homogeneous feedforward chain of

303: identical neurons such that each neuron connects via a

304: single identical excitatory synapse with the next

305: neuron of the chain. The assumption of homogeneity is

306: not realistic biologically but reduces the parameter

307: space of the calculations and increases our ability to

308: understand how the existence and properties of a

309: propagating burst depend on parameters.

310: 

311: Our 1d~model leaves out many experimental details (we

312: discuss this further in

313: Section~\ref{sec:results-discussion} below), of which

314: two are especially important, the need for multiple

315: pathways and the presence of inhibitory neurons.

316: First, as pointed out by Abeles~\cite{Abeles91} and

317: others, transmission of information along a

318: one-dimensional chain is not robust since damage to any

319: neuron in the chain can alter or stop the transmission

320: of information. (That~HVC is robust is demonstrated by

321: the fact that the adult song of a zebra finch changes

322: little over the life of the bird, even though there is

323: ongoing neurogenesis and hence a steady turnover of

324: \hvcra~neurons~\cite{NottebohmJN02,Wang02}.) If some

325: kind of propagation occurs in~HVC, there must be

326: parallel pathways.

327: % (not all of which have to be strictly feedforward~\cite{Jin05}).

328: 

329: \begin{figure}[htb]

330:   \centering

331:   \includegraphics[height=2.5in]{synfireScheme.eps}

332:   \caption{ (a) Schematic diagram of a feedforward

333:   synfire chain consisting of successive pools of

334:   neurons (vertical column of circles). Each pool

335:   contains excitatory \hvcra~neurons that fire

336:   approximately synchronously to activate neurons in

337:   the next pool. The horizontal hollow arrow on the

338:   left represents input that can initiate the synfire

339:   chain. The hollow arrows at the top of each pool

340:   indicate efferents that could convey information from

341:   a given pool to other brain areas. (b) Schematic

342:   diagram of the one-dimensional homogeneous

343:   feedforward excitatory chain that we study in this

344:   paper. In the biologically unrealistic but

345:   theoretically convenient case that all neurons in the

346:   synfire chain of~(a) are identical, that all neurons

347:   in one pool connect to neurons in the next pool in

348:   the same way with identical synapses, that axonal

349:   delays between pools can be ignored, and there is no

350:   noise, the 1d-chain will have dynamics identical to a

351:   synfire chain whose spikes have become fully

352:   synchronized. }

353:   \label{fig:syn-fire-chain}

354: \end{figure}

355: 

356: A widely studied example of a feedforward neuronal

357: architecture with multiple pathways is a synfire

358: chain~\cite{Abeles91,Diesmann99,Ikegaya04}, see

359: Fig.~\ref{fig:syn-fire-chain}(a). Although there are

360: variations of this architecture, a synfire chain is

361: usually described as a strictly feedforward chain of

362: pools of neurons such that the neurons in one pool

363: project only to neurons of the next pool. Reliable

364: transmission occurs even in the presence of noise or of

365: network heterogeneities because of the multiple paths,

366: and also because each neuron receives spikes from many

367: neurons of the previous pool and the spikes

368: spontaneously become highly synchronized as propagation

369: continues~\cite{Abeles91,Diesmann99}. The arrival of

370: synchronized spikes at the many synapses of a given

371: synfire neuron causes a large postsynaptic current that

372: triggers new spikes with a probability that can

373: approach certainty.

374: 

375: The idea that the sparse brief bursts observed in

376: \hvcra~neurons are causally connected by sequential

377: firing through some kind of chain has been proposed by

378: experimentalists~\cite{Fee04,Leonardo05} and assumed in

379: several theoretical

380: studies~\cite{Doya98,Abarbanel04jn,Berger04}. In a

381: recent experiment~\cite{Ikegaya04}, calcium imaging and

382: timing studies of intracellular recordings have

383: suggested the existence of synfire chains in the visual

384: cortices of mice and cats but no similar imaging study

385: has yet been carried out for songbirds.

386: 

387: The dynamics of our one-dimensional homogeneous chain

388: is equivalent to the dynamics of a homogeneous

389: noiseless synfire chain (homogeneity here means equal

390: numbers of identical neurons per pool all connected in

391: identical ways from one pool to the next) such that all

392: spikes within a given pool have become perfectly

393: synchronized. This is an important observation since

394: the synapse that connects one neuron to the next in the

395: 1d~chain of Fig.~\ref{fig:syn-fire-chain}(b) then has

396: to be strong to correspond to the many synapses of a

397: neuron in the full synfire chain. In general, noise,

398: heterogeneities, delay times, and imperfect

399: synchronization of spikes cause the dynamics of a

400: synfire chain to differ from that of our 1d~chain so

401: that simulations of real synfire chains are needed to

402: understand these effects on the propagation of a

403: burst~\cite{Jin05,Li06II,Li06thesis}.

404: 

405: A second important experimental detail that we leave

406: out of our model is the presence of the inhibitory

407: HVC~interneurons. Our justification for this assumption

408: is simply that the function of these interneurons is

409: not known. Proposed mathematical models for sequence

410: generation and sequence recognition incorporate

411: inhibitory neurons in various

412: ways~\cite{Kleinfeld89,Drew03,Huerta04,Abarbanel04jn}

413: but there is no support yet for these models in the

414: context of the songbird system. We argue that ignoring

415: the inhibitory neurons is a useful first step toward

416: understanding the HVC~microcircuitry since the

417: calculations discussed below show that a minimal

418: idealized 1d~chain of identical excitatory neurons

419: already suffices to produce a stable propagating burst,

420: in which case other effects such as noise,

421: inhomogeneities, and interneurons could be included

422: perturbatively. Further, some experiments suggest that

423: ignoring the interneurons may not be too severe an

424: assumption.  HVC~interneurons tend to fire rapidly,

425: chronically, and in approximate synchrony throughout a

426: motif~\cite{Fee04}, and rapidly and chronically during

427: audition of a bird's own song~\cite{Mooney00}, so that

428: the interneuron spikes might not be closely correlated

429: with the brief sparse bursts of the \hvcra~neurons. In

430: this case, the main purpose of the interneurons might

431: to prevent runaway excitation of the projection

432: neurons.

433: 

434: We conclude this section with the technical observation

435: that our assumption that the 1d~chain is homogeneous

436: and that there is no delay along axons (the axons along

437: the chain have effectively zero length) allow the

438: 1d~chain to be simulated by integrating numerically a

439: \textit{single} neuron with a single afferent

440: synapse. During the integration, the output of the

441: neuron can be stored in memory and then used as input

442: to the same neuron which then represents the next

443: neuron in a chain. A related idea was used by

444: Reyes~\cite{Reyes03}, who used a single real

445: hippocampal neuron in a slice together with the dynamic

446: clamp technique to simulate a synfire chain consisting

447: of many pools of identical neurons.

448: 

449: 

450: \subsection{A Leaky Integrate-and-Fire (LIF) Model of an \hvcra~Neuron}

451: 

452: We studied propagation of bursts along 1d~chains

453: consisting of one of two types of neurons, leaky

454: integrate-and-fire (LIF) neurons that are described in

455: this section, and more realistic conductance-based

456: Hodgkin-Huxley (HH) neurons that are described in the

457: next section. Neither neuronal model bursts

458: intrinsically when subject to a direct current (DC)

459: external stimulus. Provided that an initial burst is

460: applied to the first neuron of the 1d~chain, we find

461: that successive neurons are capable of

462: generating an identical output burst when stimulated

463: itself by a burst.

464: 

465: With respect to conductance-based models, LIF

466: models~\cite{Dayan01} are attractive since they have

467: fewer parameters, lead to efficient numerical

468: simulations, and are sometimes amenable to analytical

469: methods. The $k$th~LIF neuron of the 1d~chain ($k \ge

470: 1$) was modeled as a first-order non-autonomous

471: ordinary differential equation

472: \begin{equation}

473: \label{eq:lif}

474:   \tau_m \frac{dv^k}{dt} = v_0 - v^k  

475:    +  R \left( I^k_e(t) +  I^k_S(t) \right) ,

476: \end{equation}

477: where the parameter~$\tau_m$ is the capacitive

478: $RC$-time scale of the neuronal membrane, the

479: variable~$v^k(t)$ is the potential difference across

480: the membrane of the $k$th~neuron in the chain, the

481: external current~$I^k_e(t)$ of the $k$th~neuron is a

482: specified function (in this paper applied only to the

483: first neuron to initiate activity in the chain), the

484: synaptic current~$I^k_S(t)$ arises from spikes in the

485: previous $(k-1)$st~neuron of the chain (see

486: Eqs.~(\ref{eq:total-synaptic-current})

487: and~(\ref{eq:lif-syn-current}) below), the

488: parameter~$v_0$ is the resting potential toward which

489: the potential~$v$ asymptotes in the absence of external

490: and synaptic currents ($I_e=I_S=0$), and the

491: parameter~$R$ is the total resistance of the membrane.

492: 

493: Eq.~(\ref{eq:lif}) was supplemented with the usual

494: spiking rule that, whenever the potential~$v^k(t)$ is

495: increasing and crosses a specified threshold

496: value~$v_\textrm{thresh} > v_0$, the neuron is assumed

497: to have spiked and the potential~$v^k$ is instantly and

498: discontinuously decreased to a reset value

499: $v_\textrm{reset} < v_0$. An absolute refractory period

500: was included in this LIF~model by freezing the

501: potential~$v^k$ to the value~$v_\textrm{reset}$ for a

502: time interval $t_\textrm{refract}$ after a spike.  We

503: found that including a refractory interval of

504: $t_\textrm{refract}= 1\,\rm ms$ produced only minor

505: qualitative changes to the results obtained in the

506: absence of a refractory period, for example a 1~ms

507: refractory period expanded by a modest amount the

508: basins of attraction for the different stable

509: propagating bursts discussed in

510: Section~\ref{sec:results-discussion} (see

511: Fig.~\ref{fig:steadyPlot} below). Although

512: \hvcra~neurons show some accommodation~\cite{Mooney00},

513: we did not include this detail in our LIF~model.

514: 

515: The total postsynaptic current~$I^k_S(t)$ of the

516: $k$th~neuron in the chain was assumed to be a linear

517: accumulation

518: \begin{equation}

519:    I^k_S(t) = \sum_i I_s(t-t^{k-1}_i) ,

520:    \label{eq:total-synaptic-current}

521: \end{equation}

522: of postsynaptic currents~$I_s(t-t^{k-1}_i)$

523: associated with spikes that occurred in the previous

524: ${(k-1)}$st~neuron of the chain at times~$t^{k-1}_i <

525: t$. (Eq.~(\ref{eq:total-synaptic-current}) also assumes

526: that there are no time delays arising from the time for

527: a spike to propagate from one neuron to the next.) The

528: time dependence of the synaptic conductivity associated

529: with a single spike at time~$t^k_i=0$ was modeled as a

530: double exponential~\cite{Dayan01}, leading to the

531: following expression for the postsynaptic current per

532: spike:

533: \begin{equation}

534:    I_s(t,v) = n I_0 \left( 

535:     e^{\displaystyle -t/\tau_1} -  e^{\displaystyle

536:     -t/\tau_2}

537:     \right) .

538:    \label{eq:lif-syn-current}

539: \end{equation}

540: The slow and fast time constants~$\tau_1$ and~$\tau_2$

541: had respectively the values~1.1~ms and~0.2~ms to match

542: roughly the~4.0~ms onset-to-peak time constants

543: observed in paired intracellular recordings of

544: \hvcra~neurons~\cite{Mooney05}.  The

545: value~$I_0=0.3\,\rm nA$ was chosen to give a 1.0~mV

546: peak excitatory post-synaptic potential (EPSP) which is

547: consistent with experiment~\cite{Mooney05}.  The

548: number~$n$, which varied from~1 to~32, indicated the

549: synaptic strength in terms of the number of synchronous

550: spikes that a neuron would receive if placed in a

551: uniform synfire chain with~$n$ neurons per pool and all

552: neurons of one pool connecting to all neurons in the

553: next pool (with synapses of the same form

554: Eq.~(\ref{eq:lif-syn-current})). The range~1 to~32 was

555: found empirically to span the range of stable dynamics,

556: see Fig.~\ref{fig:steadyPlot} below.

557: 

558: An homogeneous chain was then obtained by using for

559: each LIF~neuron the same ten parameters~$\tau_m$,

560: $v_0$, $R$, $v_\textrm{thresh}$, $v_\textrm{reset}$,

561: $t_{\rm refract}$, $n$, $I_0$, $\tau_1$, and $\tau_2$,

562: and the same function Eq.~(\ref{eq:lif-syn-current})

563: for the postsynaptic current per spike. A time

564: constant~$\tau_m = 15\,\rm ms$ was used to approximate

565: the {\it in vivo} response time~\cite{Mooney00}, and a

566: membrane resistance of~$R=60\,\rm M\Omega$ was chosen

567: to match the impedance of

568: \hvcra~neurons~\cite{Mooney00}.  Other parameter values

569: used were~$v_0=-70$~mV, $v_\textrm{thresh}=-55$~mV,

570: $v_\textrm{reset}=-75$~mV, and~ $t_{\rm

571: refract}=1.0$~ms.

572: 

573: The chain of LIF~neurons represented by

574: Eqs.~(\ref{eq:lif})-(\ref{eq:lif-syn-current}) was

575: integrated by using a forward-Euler

576: method~\cite{Kincaid96} with a constant time step

577: of~$\Delta{t}=0.01\,\rm ms$.  (Smaller time steps by

578: factors of four did not lead to significant changes in

579: the results.) The supplementary reset and refractory

580: rules were applied at the end of each time step. For

581: most runs, a zero external current~$I^k_e=0$ was

582: assumed for each neuron of the chain except for the

583: first neuron, for which a step function was used to

584: stimulate a burst.  The code was programmed and run

585: using the computational mathematics program Matlab

586: version~6.5~\cite{Matlab}.

587: 

588: 

589: \subsection{A  Single-Compartment Hodgkin-Huxley

590:   Model of an \hvcra~Neuron}

591: 

592: \label{sec:hh-model-of-hvcra-neuron}

593: 

594: The second neuronal model that we used to study the

595: propagation of a burst in a 1d~chain of excitatory

596: neurons was a single-compartment model based on the

597: Hodgkin-Huxley equations with five representative

598: conductances.  The evolution equation for the membrane

599: potential~$v^k(t)$ of the~$k$th neuron in the chain was

600: \begin{equation}

601:   C_m {dv^k \over dt} = \sum_{i=1}^5 g_i(t,v^k)

602:    \left( v_i - v^k \right)  +  I^k_e(t) + I^k_S(t) .

603:   \label{eq:HH-model}

604: \end{equation}

605: The symbols have the following meanings: $C_m$ is the

606: total membrane capacitance, $g_i(t,v)$~is the

607: voltage-dependent conductance of the $i$th~kind of

608: membrane channel, $v_i$~is the resting potential for

609: the $i$th~channel, and the currents~$I^k_e(t)$

610: and~$I^k_S(t)$ have the same meaning as in

611: Eq.~(\ref{eq:lif}). Appendix~\ref{appendix:equations-parameters}

612: gives the details of parameter values and other

613: evolution equations related to the conductances~$g_i$.

614: 

615: Although more realistic in terms of its time

616: dependence, the HH~model Eq.~(\ref{eq:HH-model}) is not

617: necessarily more scientifically appropriate than

618: LIF~models since the properties of the various membrane

619: conductances are only partially known for

620: HVC~neurons~\cite{Kubota91,Kubota98,Dutar98,Mooney00},

621: the spatial distribution of the channels in HVC~neurons

622: has not been determined (which would be needed to

623: construct accurate multi-compartment models), and

624: little is known about the type, strengths, and

625: locations of synapses in~HVC. We did find that a

626: single-compartment HH~neuron with five conductances is

627: only able to match some features of the

628: experimental~HVC data, and it is not clear which

629: experimental details are more important than others to

630: include in the process of fitting a mathematical model

631: to data. For these reasons, and because this paper is

632: concerned whether in principle an HVC-like burst can

633: propagate stably through an excitatory chain, the

634: conductances of the HH~model Eq.~(\ref{eq:HH-model})

635: were only loosely based on existing HVC~data.

636: 

637: Our starting point for choosing the conductances in

638: Eq.~(\ref{eq:HH-model}) was a recent paper by Prinz et

639: al~\cite{Prinz03}, which showed that an

640: eight-dimensional phase space obtained by varying the

641: maximum conductances of eight channels (whose

642: functional properties were obtained from lobster

643: stomatogastric neuronal data) contained a great

644: diversity of dynamical behavior. We chose the

645: functional forms of the leakage, sodium, and potassium

646: channels from this paper and added two other channels

647: as suggested by current-clamp data of~HVC

648: neurons~\cite{Kubota91}: a fourth channel was a

649: low-threshold transient calcium current, and a fifth

650: channel was a calcium-activated potassium current

651: (denoted in the following by the symbol~KCa).  The

652: calcium channel was chosen to activate around -50~mV

653: and inactivate around -80~mV with time scales such that

654: the channel could be activated

655: transiently~\cite{Kubota91,Huguenard92}.

656: %%% mrli: would help to explain what is meant here by

657: %%% different time scales

658: The KCa~conductance activates the potassium current

659: after intracellular calcium concentration rises, with a

660: functional form adopted from Yamada's

661: model~\cite{Yamada98}.  The maximum conductances of all

662: five conductances were adjusted iteratively by hand to

663: produce a short spike width and to generate a spike

664: adaptation with a high initial firing rate as observed

665: by Kubota~\cite{Kubota91}.

666: %%% Please be more specific: how short was the spike

667: %%% width, what was the value of the high initial firing rate?

668: A reasonably good fit to the Mooney {\it in~vivo} data

669: (see Fig.~1 of Ref.~\cite{Mooney00}) unfortunately

670: could not also match Kubota's {\it in~vitro} spike

671: profiles of Ref~\cite{Kubota98} and vice versa. (We

672: hope to obtain more complete fits to HVC~data by adding

673: more channels and more compartments when more

674: anatomical and physiological data become available.)

675: 

676: %%% MengRu: please add further details of what you

677: %%% actually did. These details should especially

678: %%% appear in your thesis.

679: 

680: The total synaptic current was also assumed to satisfy

681: the linear relation

682: Eq.~(\ref{eq:total-synaptic-current}) but an

683: alpha~function~\cite{Dayan01} was used instead of

684: Eq.~(\ref{eq:lif-syn-current}) to approximate the

685: time-dependent probability of a channel opening.  The

686: excitatory postsynaptic current (EPSC) arising from a

687: spike that occurs at time~$t=0$ had the form

688: \begin{equation}

689:    I_s(t;v) = 

690:     g_s C \left( t \over \tau_s \right) e^{\displaystyle -t/\tau_s} 

691:    \left( v_s - v \right) ,

692:   \label{eq:hh-syn-current}

693: \end{equation}

694: where the constant~$C$ with value~$e$ normalizes the

695: maximum value of the expression

696: $C(t/\tau_s)\exp(-t/\tau_s)$ to~1. This normalization

697: makes the parameter~$g_s$ the maximum conductance,

698: which we varied over the range 0 to 0.55~nS.  In

699: Eq.~(\ref{eq:hh-syn-current}), the synaptic time

700: constant~$\tau_s$ was fixed with value~7~ms and

701: %%% mrli: where did the value for tau_s come from?

702: %%% Should give reference if you have one

703: the synaptic reversal voltage~$v_s$ was set to~120~mV

704: to make the synapse excitatory.

705: Eqs.~(\ref{eq:HH-model}),

706: (\ref{eq:total-synaptic-current})

707: and~(\ref{eq:hh-syn-current}) together with the

708: evolution equations for gate variables given in

709: Appendix~\ref{appendix:equations-parameters} were

710: integrated with the Neuron simulation code

711: version~5.6~\cite{NeuronCode}, with a constant time step

712: of~$\Delta{t}=0.1\,\rm ms$.  

713: 

714: %Another simulation with

715: %$\tau_s=1.5\,\rm ms$, $v_s = 0\,\rm mV$, $g_s$ between

716: %0.4~nS and 12~nS, and time step 0.02~ms showed the same

717: %result with more compact bursts.

718: 

719: %%% MengRu: how did you figure out that this time step

720: %%% was sufficiently accurate? What is the importance

721: %%% of mentioning that the bursts were more compact?

722: 

723: %%% What value of synaptic reversal potential v_s did

724: %%% we use for HH synapses? What other details do we 

725: %%% need to add?

726: 

727: 

728: \section{Results and Discussion}

729: 

730: \label{sec:results-discussion}

731: 

732: 

733: \subsection{Results for a 1d Homogeneous Chain of HH~Neurons}

734: 

735: The calculations discussed below of one-dimensional

736: homogeneous chains using the LIF or HH~neurons of the

737: previous section with excitatory synapses show that a

738: brief high-frequency burst similar to that observed in

739: \hvcra~neurons during singing~\cite{Hahnloser02} can

740: propagate stably over a range of parameter values.

741: These results make more plausible the hypothesis that

742: an excitatory chain of \hvcra~neurons generates the

743: sparse precisely-aligned bursts observed

744: experimentally~\cite{Hahnloser02}. Our results show

745: further that different asymptotic attractors can exist

746: for given parameter values so that hysteresis can occur

747: (different initial conditions can lead to different

748: non-transient dynamics).  The fact that the bursts

749: exist as an attractor means that there is a transient

750: time during which properties of the burst such as the

751: number of spikes and burst width (alternatively, the

752: average burst frequency) evolve until the final stable

753: values are obtained. This transient time varies with

754: the initial state used to start the chain and with the

755: choice of parameters.

756: 

757: Since these qualitative conclusions turned out to be

758: similar for LIF and HH~neuronal models, we report here

759: results mainly for the HH~models of

760: Section~\ref{sec:hh-model-of-hvcra-neuron} and mention

761: briefly in the next section how the results differ for

762: a chain of~LIF neurons. Since the parameter spaces for

763: LIF and HH~neurons are high-dimensional (10~parameters

764: for~LIF, about 25~parameters for the HH~neurons, and

765: these do not include the choice of the functional

766: form~$I_s(t,v)$ for the~EPSC per spike), to establish

767: our key results we varied only one parameter

768: systematically, namely the maximum synaptic

769: conductance~$g_s$ of Eq.~(\ref{eq:hh-syn-current}) for

770: HH~neurons, or equivalently the number of synchronous

771: presynaptic spikes~$n$ of

772: Eq.~(\ref{eq:lif-syn-current}) for LIF neurons.

773: 

774: \begin{figure}[tb]

775:   \includegraphics[height=2.7in]{burstPropagate.eps}

776:   \caption{ Propagation of an initial 5-spike burst

777:     through an homogeneous one-dimensional excitatory

778:     chain of HH~neurons, for different maximum synaptic

779:     conductances~$g_s$ of

780:     Eq.~(\ref{eq:hh-syn-current}). Each column is

781:     corresponds to a fixed~$g_s$ value, and successive

782:     rows of a given column show the membrane

783:     potential~$v^k(t)$ as a function of time for

784:     successive neurons in the chain ($k=1, 2,

785:     \cdots$). {\bf (a)}~For a weak synaptic coupling

786:     strength~$g_s=0.1\,\rm nS$, the initial burst

787:     decays by the third neuron.  {\bf (b)}~For a

788:     stronger coupling~$g_s=0.2\,\rm nS$, the initial

789:     bursts decays to an invariant single spike.  {\bf

790:     (c)}~For~$g_s=0.3\,\rm nS$, the initial burst

791:     evolves to a stable state with three unevenly

792:     spaced spikes.  {\bf (d)}~For still stronger

793:     couplings $g_s \ge 0.45\,\rm nS$, the initial burst

794:     can be unstable for some initial states and the

795:     number of spikes increases steadily.}

796:   \label{fig:burst-propagation}

797: \end{figure}

798: 

799: Figure~\ref{fig:burst-propagation} shows how the same

800: initial burst of five spikes propagates through a

801: 1d~excitatory chain of HH~neurons for several different

802: values of the maximum synaptic conductance~$g_s$. The

803: initial burst was created by using an external

804: current~$I^1_e(t)$ of Eq.~(\ref{eq:HH-model}) for the

805: first neuron in the chain to inject a square pulse of

806: current that caused five spikes to appear in rapid

807: succession.  For a synaptic coupling strength~$g_s$

808: smaller than about~$0.1\,\rm nS$,

809: Fig.~\ref{fig:burst-propagation}a shows that the

810: initial burst rapidly dies out by the third neuron of

811: the chain and all spikes disappear. Over a range of

812: stronger couplings $0.2 < g_s < 0.4\,\rm nS$, the

813: initial bursts evolves into a stable invariant

814: propagating burst that can have one to five spikes

815: (columns (b) and~(c) of

816: Fig.~\ref{fig:burst-propagation}). For stronger

817: couplings $g_s > 0.4\,\rm nS$, the initial burst is

818: unstable and the number of spikes grows steadily

819: without limit. However, other initial conditions can

820: lead to stable bursts in this range, an example of

821: hysteresis.

822: 

823: \begin{figure}[tb]

824:   \centering

825:   \includegraphics[height=2.7in]{initCondProp2.eps}

826:   \caption{(Color onlin) The dependence of transient dynamics and

827:     resulting attractors of propagating bursts on the

828:     initial injecting current magnitude~$I^1_e(t)$ and

829:     on the synaptic conductance~$g_s$. For weak

830:     couplings with $g_s < 0.2\,\rm nS$, all initial

831:     states decay to zero (not shown here). {\bf (a)}

832:     For~$g_s=0.3\,\rm nS$, there are multiple basins of

833:     attraction corresponding to asymptotic propagating

834:     bursts with 1, 2, or~3 spikes. Transients decay

835:     rapidly by the third neuron of the chain. {\bf (b)}

836:     Increasing the synaptic strength to~$g_s =0.35\,\rm

837:     nS$ increases the number of attractors, with final

838:     bursts containing 1 to~5~spikes.  {\bf (c)} For a

839:     larger synaptic conductance~$g_s=0.45\,\rm nS$,

840:     initial conditions lead to stable or unstable

841:     states. An initial burst with 6~spikes will evolve

842:     to a slightly different invariant burst with also

843:     6~spikes (squares). }

844:   \label{fig:initCondProp}

845: \end{figure}

846: 

847: Fig.~\ref{fig:initCondProp} provides a more global

848: understanding of the transient dynamics and resulting

849: attractors. In each panel, the vertical axis indicates

850: the number of spikes observed in a burst while the

851: horizontal axis indicates the position~$k$ of a neuron

852: along the chain. For a weak synaptic coupling~$g_s<

853: 0.2\,\rm nS$, all initial states decay to zero spikes

854: (not shown in the figure).  For a synaptic coupling

855: of~$g_s=0.3\,\rm nS$ (Fig.~\ref{fig:initCondProp}a),

856: initial bursts with three or more spikes decay within

857: four neurons to a common final burst of three

858: spikes. An initial burst with two spikes evolves

859: slightly to an invariant burst with also two spikes,

860: and a similar conclusion holds for an initial burst

861: with a single spike. As the synaptic strength is

862: increased to~$g_s=0.35\,\rm nS$

863: (Fig.~\ref{fig:initCondProp}b), the number of

864: attractors increases so that stable bursts with one to

865: five spikes are observed depending on the initial state

866: of the first neuron. The transient time is still rather

867: short, with the final number of spikes stabilizing in

868: all cases within three neurons ($k \le 3$). Finally for

869: synaptic couplings stronger than about~$0.4\,\rm nS$

870: (Fig.~\ref{fig:initCondProp}c), stable and unstable

871: states are observed depending on the initial state of

872: the first neuron. The unstable bursts all grow at the

873: same rate, with an extra spike appearing for each next

874: neuron traversed.

875: 

876: \begin{figure}[tb]

877:   \centering

878:   \includegraphics[height=2.7in]{twoDMap2.eps}

879:   \caption{ (Color online) Different synaptic conductances~$g_s$ lead

880:     to different final propagating burst widths and

881:     spike numbers, which shows that the total burst

882:     duration is another important characteristic of the

883:     final attractor.  For $g_s = 0.3\,\rm nS$, the

884:     curve of block symbols evolves into the point~E1.

885:     For $g_s = 0.35 \,\rm nS$, the curve with diamond

886:     symbols ends in the point~E2.  For $g_s = 0.45\,\rm

887:     nS$, the propagation is unstable (upper triangle)

888:     except for the initial condition at point~B2.  }

889:   \label{fig:transient-states}

890: \end{figure}

891: 

892: The transient dynamics of Fig.~\ref{fig:initCondProp}

893: are presented in somewhat finer detail in

894: Fig.~\ref{fig:transient-states}, which shows that a

895: non-transient propagating bursts is characterized by at

896: least two parameters, their total width and the number

897: of spikes.  For example, the same beginning state~$B_1$

898: can evolve into three different end points depending on

899: the synaptic coupling strength~$g_s$: point~$E_1$ with

900: three spikes, point~$E_2$ with 4~spikes and with a

901: slightly longer width of~45~ms, and an unstable state

902: (triangles).

903: 

904: \begin{figure}[tb]

905:   \centering

906:   \includegraphics[height=2.7in]{steadyPlot.eps}

907:   \caption{ Plot of the number of spikes observed in a

908:   final non-transient propagating burst for different

909:   initial amplitudes of a square current pulse of fixed

910:   duration~20 ms and for different maximum synaptic

911:   conductances~$g_s$.  Different basins of attraction

912:   are observed, with 0-spike and unstable bursts having

913:   the largest basins, and roughly equal size basins for

914:   bursts with 1 to~5 spikes in the final burst.}

915:   \label{fig:steadyPlot}

916: \end{figure}

917: 

918: Finally, Fig.~\ref{fig:steadyPlot} indicates the

919: approximate basins of attraction that exist for

920: different initial current amplitudes (horizontal axis)

921: and different maximum synaptic conductances (vertical

922: axis). The plotted numbers indicate the number of

923: spikes observed in the asymptotic state for the

924: specified axis values.  This plot shows again how all

925: states decay away for sufficiently weak synaptic

926: coupling, all states are unstable for sufficiently

927: strong coupling, and that stable propagating bursts

928: with~1 to~6 spikes can be found for intermediate

929: coupling strengths. The basins of attraction turn out

930: to be of comparable size except for the tiny basin

931: corresponding to a 6-spike burst.

932: 

933: 

934: \subsection{Results for a 1d Homogeneous Chain of LIF

935: Neurons}

936: 

937: \label{sec:results-1d-homog-lif-chain}

938: 

939: Our results for homogeneous chains of~LIF neurons were

940: qualitatively similar to the above results for chains

941: of HH~neurons and we do not show the corresponding

942: results.  Instead of initiating the chain with a

943: current pulse, the first LIF~neuron of the chain was

944: initiated with a synaptic current corresponding to a

945: burst of equally-spaced spikes whose times of

946: occurrences~$t^0_i$ were specified as initial data (see

947: Eq.~(\ref{eq:total-synaptic-current})).  Similarly,

948: instead of a maximum conductance~$g_s$ being varied,

949: the synaptic coupling strength~$n$ of

950: Eq.~(\ref{eq:lif-syn-current}) was varied between~1

951: and~32 which we found to span the same kinds of states

952: discussed in Figs.~\ref{fig:burst-propagation}

953: and~~\ref{fig:steadyPlot}, from a 0-spike final state

954: to an initial burst that grew without bound in width

955: and in the number of spikes.  Because the LIF neurons

956: have no intracellular calcium dynamics, the number of

957: spikes in the final state was primarily determined by

958: the initial spike number. In contrast, the

959: intracellular calcium current of a HH neuron can boost

960: depolarization and maintain propagating bursts in

961: cohesion so that chains of HH~neurons have bigger

962: basins of attraction for stable bursts.

963: 

964: % When LIF synfire chains have no refractory period, the

965: % basins of attraction are narrower and the propagation 

966: % tends to amplify to become unstable.  The refractory period

967: % plays a role in shutting down excessive synaptic 

968: % current and in stabilizing a synfire chain.

969: 

970: 

971: \subsection{Other Hypotheses and Related Theory}

972: \label{sec:other-hypothese-and-related-theory}

973: 

974: We discuss briefly here some other hypotheses and

975: related theoretical work that might explain the data of

976: Hahnloser et al~\cite{Hahnloser02} and of

977: Mooney~\cite{Mooney00} but that do not involve the

978: propagation of bursts through a feedforward network.

979: 

980: While a feedforward network, especially a synfire

981: chain, is one of the simplest concepts that might

982: explain sparse precisely-timed bursting, there is much

983: theoretical work dating back to the~1980s which shows

984: that a recursive network with inhibitory and excitatory

985: connections is capable of learning and producing

986: different kinds of temporal

987: sequences~\cite{Kleinfeld89,Hertz91}. These networks

988: generalize the Hopfield attractor model of associative

989: memory~\cite{Hopfield82,Hertz91} by allowing

990: non-symmetric couplings between pairs of neurons, and

991: by using two kinds of synapses, ``fast'' synapses that

992: stabilize a given network state, and ``slow'' synapses

993: that cause successive transitions between the

994: quasistatic network states.  Thus there is no

995: difficulty in principle for a recursive network to

996: store, generate, or learn many temporal sequences of

997: the sort observed in~HVC. (We note that other classes

998: of recursive models are possible, e.g., Huerta and

999: Rabinovich~\cite{Huerta04} discuss a model with neurons

1000: that are strictly inhibitory or strictly excitatory,

1001: rather than allow any neuron to have inhibitory and

1002: excitatory connections with other neurons.)

1003:  

1004: Asymmetric Hopfield-like recursive networks have

1005: several attractive features for modeling

1006: HVC~dynamics. The HVC inhibitory

1007: neurons~\cite{Mooney00} can be incorporated in a

1008: natural way since temporal sequences are stored via

1009: neuronal connection strengths~$J_{ij}$ that typically

1010: have positive (excitatory) and negative (inhibitory)

1011: values. Recursive networks that generate sequences can

1012: evolve via a simple Hebbian learning

1013: rule~\cite{Kleinfeld89} from densely interconnected

1014: neurons (see Figs.~2 and~3 of Ref.~\cite{Mooney00}) and

1015: so are possibly easier to form during maturation of

1016: brain tissue than purely feedforward

1017: networks. Asymmetric Hopfield models are naturally

1018: fault-tolerant and error-correcting so that details of

1019: a stored sequence are weakly affected if neurons or

1020: synapses are modified, deleted, or added. Finally, a

1021: single network of this kind is capable of storing and

1022: generating many different temporal sequences, which is

1023: consistent with the ability of some songbirds to learn

1024: and sing many different songs~\cite{Catchpole95}.

1025: 

1026: However, further work is needed to determine whether

1027: existing recursive models~\cite{Kleinfeld89} are

1028: capable of taking into account specific experimental

1029: details of~HVC such as the fact that an \hvcra~neuron

1030: fires briefly just once during a one-second-long motif

1031: (a neuron is more likely to fire multiple times during

1032: a motif if there are recurrent pathways), that time

1033: interval between successive bursts is short (estimated

1034: to be about 10~ms~\cite{Fee04}, this is possibly too

1035: short for a separation of time scales to exist between

1036: slow and fast synapses~\cite{Kleinfeld89}), that there

1037: is a precise alignment of bursts with auditory features

1038: during singing~\cite{Hahnloser02} and during

1039: audition~\cite{Mooney00}, and especially that

1040: HVC~inhibitory neurons fire tonically while the

1041: HVC~projection neurons fire sparsely during

1042: singing~\cite{Hahnloser02} and during

1043: audition~\cite{Mooney00}. (There is no dynamical

1044: distinction between inhibitory and excitatory synapses

1045: in the recursive models~\cite{Kleinfeld89}.)

1046: 

1047: Another alternative explanation for the origin and

1048: precise alignment of \hvcra~bursts during singing and

1049: during audition is that these bursts are driven by

1050: precisely timed external inputs. This possibility is

1051: suggested by the fact that other nuclei such as~NIF

1052: and~Uva are known to provide distributed input to HVC,

1053: and to all three classes of

1054: HVC~neurons~\cite{Coleman04,Rosen06}. However, some

1055: experiments suggest that this is a less plausible

1056: explanation than intrinsic generation of bursts

1057: within~HVC. For example, preventing auditory input

1058: to~HVC by destroying a bird's cochlea (which deafens

1059: the bird) or by lesioning nucleus~NIF does not prevent

1060: a bird from singing, nor does it cause a bird's song to

1061: change substantially over time scales of days.

1062: %%% I need to get references for these points!

1063: Although extracellular recordings of an awake singing

1064: bird of the kind reported by Hahnloser et

1065: al~\cite{Hahnloser02} have not been carried out after

1066: deafening or lesioning, the fact that the song does not

1067: change immediately suggests that the same pattern of

1068: sparse bursts should still be observed

1069: in~\hvcra~neurons. Since the inputs to~HVC remain

1070: incompletely characterized, external driving, or some

1071: combination of external driving and intrinsic HVC

1072: circuitry, might be able to explain the observed bursts

1073: of \hvcra~neurons during singing.

1074: 

1075: A variation of the idea of precise external input

1076: driving the bursts in~HVC is the possibility that part

1077: of~HVC acts as a central pattern generator (CPG) that

1078: drives the \hvcra~neurons. For example, the

1079: time-ordering of bursts in \hvcra~neurons observed by

1080: Hahnloser et al~\cite{Hahnloser02} could arise from

1081: a~CPG composed of other HVC~neurons such that the~CPG

1082: connects to the \hvcra~neurons via axons of different

1083: lengths and so with different delay times. (The

1084: dynamics of the~CPG itself could be explained as a

1085: non-symmetric Hopfield model~\cite{Kleinfeld89} or a

1086: cyclic synfire chain.) A CPG-based model would differ

1087: from a synfire chain in that a given burst does not

1088: trigger causally the next burst like dominoes falling

1089: over along some path. For example, destroying all

1090: \hvcra~neurons that burst at a certain time in a

1091: CPG~model would not stop the firing of \hvcra~neurons

1092: that normally would burst a short time later. A recent

1093: \textit{in vitro} slice experiment~\cite{Solis05}

1094: suggests that HVC~has intrinsic oscillatory dynamics

1095: consistent with the existence of a~CPG, but it is not

1096: yet known how these slice data relate to the dynamics

1097: of \hvcra~neurons during song.

1098: 

1099: 

1100: \section{Conclusions}

1101: \label{sec:conclusions}

1102: 

1103: For an idealized one-dimensional homogeneous

1104: feedforward chain of excitatory non-bursting neurons,

1105: we have shown by numerical calculations that a brief

1106: high-frequency burst of two to six spikes can propagate

1107: in a stable way for various choices of parameter

1108: values. Previous studies of excitatory networks have

1109: not addressed the propagation of bursts and so have not

1110: been directly relevant for recent experiments

1111: concerning the properties of nucleus~HVC. Previous

1112: studies~\cite{Diesmann99} have also mainly involved

1113: networks of LIF~neurons (to reduce the computational

1114: effort) and so have not taken into account the more

1115: complex dynamics and realistic time scales of

1116: HH~neurons, especially the effects of a calcium current

1117: which experiments have shown to be present in

1118: HVC~neurons~\cite{Kubota91,Kubota98}. Our results show

1119: that stable bursts exist over a range of parameter

1120: values for both LIF and~HH neuronal models so that the

1121: propagation of stable bursts in a homogeneous chain is

1122: not sensitive to the choice of model or of model

1123: parameters.

1124: 

1125: These results make plausible the hypothesis that the

1126: sparse precisely-aligned high-frequency brief bursts

1127: observed by Hahnloser et al~\cite{Hahnloser02} in

1128: \hvcra~neurons during singing are intrinsic (do not

1129: need external input) and are causally connected in that

1130: one neuron initiates bursting in the next neuron and so

1131: on until the end of the chain. However, other

1132: explanations of the sparse precise bursts are possible

1133: as discussed in

1134: Section~\ref{sec:other-hypothese-and-related-theory},

1135: and further experiments, especially using methods that

1136: can analyze many neurons at once with good time

1137: resolution, would be valuable in helping to distinguish

1138: feedforward chains from external driving, recursive

1139: networks, or other mechanisms.

1140: 

1141: We have not addressed in this paper why \hvcra~neurons

1142: burst in the first place in, i.e., why information is

1143: transmitted as bursts rather than as a population of

1144: spikes. (Sparse firing can aid the learning of

1145: sequences as pointed out by Fiete et al~\cite{Fiete04}

1146: but this does not require bursts specifically.) There

1147: is somewhat of a paradox here in that the commonly

1148: posited purpose of a burst is to increase the

1149: likelihood of accurate transmission through a noisy

1150: network~\cite{Lisman97,Prida97}, while a similar goal

1151: is achieved without bursts by using a synfire

1152: chain~\cite{Abeles91}. Thus if a synfire chain is the

1153: correct architecture for \hvcra~neurons, it is less

1154: clear why bursts are needed, especially since HVC~has

1155: much less noise than cortical neurons and HVC~synapses

1156: are substantially stronger than cortical

1157: synapses~\cite{Mooney05}. Unpublished calculations by

1158: the authors~\cite{Li06II,Li06thesis} that extend the

1159: present paper to noisy heterogeneous synfire chains do

1160: suggest that bursts are important for achieving robust

1161: propagation even for a synfire architecture.

1162: 

1163: We finish this paper by summarizing some consequences

1164: of the chain hypothesis and of our calculations for

1165: recent and future songbird experiments.

1166: \begin{enumerate}

1167:   

1168: \item A simple consequence of the feedforward nature of

1169:   an excitatory chain is that an experimentalist in

1170:   principle should be able to initiate singing of an

1171:   arbitrary contiguous segment of a motif. For example,

1172:   if it is possible to stimulate an intermediate part

1173:   of a chain in such a way that a burst begins to

1174:   propagate toward the end of the chain, a songbird

1175:   might then sing a motif that starts somewhere in the

1176:   middle and that then continues to its end. Similarly,

1177:   physically terminating the chain at some intermediate

1178:   point (say with a local lesion) could cause a motif

1179:   to terminate before its usual endpoint, or some

1180:   combination of stimulating in the middle of the chain

1181:   and lesioning later in the chain could be implemented

1182:   in which case the motif could start somewhere in the

1183:   middle and terminate prematurely.

1184: 

1185:   Electrical stimulation of~HVC via a single

1186:   extracellular electrode in an awake behaving bird

1187:   does not initiate singing~\cite{Vu94,Vicario95,Vu98}

1188:   but instead resets a motif to its beginning (if the

1189:   stimulus occurs while a bird was singing). The

1190:   failure to initiate the singing of a motif, or part

1191:   of a motif, does not rule out the existence of an

1192:   excitatory chain but instead could imply that a

1193:   carefully-arranged pattern of input spikes might be

1194:   needed to initiate activity at some point in the

1195:   chain, especially if the network is a synfire chain

1196:   for which most neurons in a pool must fire in near

1197:   synchrony for neurons in the next pool to

1198:   fire. Identifying the location of \hvcra~neurons in a

1199:   chain by, say, optical imaging would be a valuable

1200:   prior step that could suggest how to work out an

1201:   electrical or photo-uncaging

1202:   protocol~\cite{Losonczy06} that could initiate a

1203:   chain at an arbitrary point along its length, or

1204:   could suggest how to lesion neurons that would

1205:   terminate propagation of a burst hence a motif

1206:   prematurely.

1207: 

1208: \item Our calculations suggest that a homogeneous chain

1209:   based on~LIF or HH~neuronal models should generally

1210:   have multiple basins of attraction.  Different kinds

1211:   of asymptotic stable bursts differing in the number

1212:   of spikes and in their total duration should be

1213:   observed in nucleus~HVC for different values of

1214:   synaptic strength and other neuronal parameters.

1215:   Further, hysteresis can be expected such that, for

1216:   identical experimental conditions, different initial

1217:   states will evolve to different kinds of asymptotic

1218:   bursts.

1219:   

1220: %   Verifying the existence of different basins of

1221: %   attraction is difficult and would require varying

1222: %   somehow the properties of HVC neurons or of their

1223: %   synapses and observing how a given initial burst

1224: %   evolves. Some properties can be altered by fatiguing

1225: %   synapses (say by repeated stimulation), by locally

1226: %   administering chemical agents (e.g., low doses of

1227: %   synaptic inhibitors to weaken synaptic transmission),

1228: %   by local photo-uncaging of neurotransmitters (which

1229: %   could strengthen synapses just before a burst

1230: %   arrives), or in the longer term, by genetic methods

1231: %   that, say, increase or decrease the amount of

1232: %   neurotransmitter stored in \hvcra\ synaptic vesicles

1233: %   or that alter the density of postsynaptic receptors.

1234: 

1235:   Hysteresis under fixed experimental conditions can be

1236:   explored by initiating different kinds of bursts.

1237:   This may be possible if the neurons at the beginning

1238:   of the chain can be identified and an appropriate

1239:   stimulation protocol worked out (see points~1 and~3).

1240: 

1241:   Most nonlinear network models, including asymmetric

1242:   Hopfield models and CPGs, also have multiple basins

1243:   of attraction so this is not a distinguishing

1244:   property of chain models. However, the number and

1245:   types of basins of attraction, and how these basins

1246:   vary under the influence of pharmaceuticals that alter

1247:   synaptic strengths, may lead to predictions that

1248:   distinguish feedforward from recursive models.

1249:   

1250: \item Chains of neurons, like the neuronal models that

1251:   make up the chains, are driven dissipative systems so

1252:   that initial states will generally evolve through a

1253:   transient before settling into an asymptotic

1254:   behavior~\cite{Berge84}.  Properties of a burst such

1255:   as its number of spikes and its set of interspike

1256:   intervals will evolve over time, with the transient

1257:   time depending on the choice of initial condition

1258:   (how far the initial state is from the attractor) and

1259:   on neuronal parameters.  Our results such as

1260:   Fig.~\ref{fig:initCondProp} for an homogeneous

1261:   HH~neuronal model suggest that transient times can in

1262:   fact be rapid, just four or fewer successive neurons.

1263:   Fig.~2 of Hahnloser et al~\cite{Hahnloser02} does not

1264:   show evidence of transient behavior since later

1265:   bursts do not seem to be statistically different from

1266:   earlier bursts during the same motif. However, only a

1267:   small number (about~20) of~\hvcra~neurons have been

1268:   sampled to date in singing birds, and neurons that

1269:   burst at the beginning of a motif have not been found

1270:   so further study involving more neurons would be

1271:   worthwhile. It may also be the case that the songbird

1272:   brain has evolved in such a way as to reduce or

1273:   eliminate the duration of transients, for example by

1274:   starting a chain with a state close to the asymptotic

1275:   burst, or by using inhibitory neurons to accelerate

1276:   convergence to synchronized firing as bursts

1277:   propagate from projection neuron to projection

1278:   neuron~\cite{Kopell04}.

1279:   

1280: \item A one dimensional chain is not robust since

1281:   propagation terminates if any neuron in the chain

1282:   dies.  A more realistic synfire chain must have at

1283:   least two neurons in a pool to be robust, and the

1284:   neurons in a pool must synchronize their spikes to

1285:   some extent to achieve reliable transmission from

1286:   pool to pool~\cite{Abeles91}. A synfire chain

1287:   hypothesis thus predicts that there must be more than

1288:   one \hvcra~neuron bursting at any give time (these

1289:   are the neurons that belong to the same pool) and

1290:   further that neurons that burst at about the same

1291:   time should have nearly synchronous bursts. Thus

1292:   Fig~2 of Ref.~\cite{Hahnloser02} may be incomplete

1293:   and a high-resolution optical study of~HVC (or

1294:   possibly a many-electrode study) during the singing

1295:   of a motif or during the audition of a bird's own

1296:   song~\cite{Mooney00} may reveal multiple

1297:   \hvcra~neurons firing bursts in synchrony.

1298:   

1299:   Abeles has pointed out that there is an inverse

1300:   relation between the number of neurons needed in a

1301:   synfire pool for reliable propagation and the average

1302:   strength of the synapses between

1303:   pools~\cite{Abeles91}. Using paired intracellular

1304:   recordings, Mooney and Prather~\cite{Mooney05} have

1305:   shown that synapses between \hvcra~neurons are

1306:   stronger than mammalian cortical neurons by about an

1307:   order of magnitude (the average~EPSP magnitude is

1308:   about 2~mV compared to about 0.2~mV in cortex) so

1309:   chains in~HVC would be expected to have about an

1310:   order of magnitude fewer neurons in a pool. Given

1311:   Abeles's estimate of about 50-100~neurons in a

1312:   cortical pool, perhaps ten or fewer neurons might be

1313:   needed in an HVC~synfire pool.  Assuming that the

1314:   HVC~bursts of average duration 6~ms are

1315:   non-overlapping, a zebra finch motif of duration

1316:   0.6~s would require about $0.6/0.006 \approx 100$

1317:   successive pools to span the motif, so a synfire

1318:   chain in~HVC might involve about $100 \times 10

1319:   \approx 1,000$ neurons. This is a much smaller number

1320:   than the 40,000 \hvcra~neurons estimated to exist

1321:   in~HVC. This suggests that a synfire chain might be

1322:   too simplistic an architecture to justify the large

1323:   number of \hvcra~neurons.

1324: 

1325: %  A more detailed understanding of the anatomy of

1326: %  HVC~synapses (numbers, kinds, spatial distributions,

1327: %  and strengths) between \hvcra~neurons would greatly

1328: %  help in our

1329: 

1330:   

1331: \item 

1332:   It is worth noting that a synfire chain architecture

1333:   can explain how an overall precise timing can be

1334:   maintained in~HVC even though there is a steady and

1335:   substantial turnover of neurons in~HVC throughout the

1336:   life of the songbird~\cite{NottebohmJN02,Wang02}.

1337:   Since perfect synchronization of a pool is not needed

1338:   to guarantee transmission of information through the

1339:   next pool, a fraction of neurons in a given pool can

1340:   alter their properties or fail without degrading the

1341:   transmission of information or its timing.

1342:   

1343: \end{enumerate}

1344: 

1345: In conclusion, our calculations support the hypothesis

1346: that the sparse bursts observed experimentally in

1347: \hvcra~neurons during singing can be understood as the

1348: propagation of bursts through an excitatory feedforward

1349: synfire chain. However, further experiments and

1350: computational studies are needed to confirm this

1351: hypothesis and to rule out the competing hypotheses

1352: discussed in

1353: Section~\ref{sec:other-hypothese-and-related-theory}.

1354: Especially interesting in the near term would be to

1355: understand some of the quantitative details of the

1356: experiments, for example the production of a brief 6~ms

1357: burst of four spikes corresponding to a frequency of

1358: about 600~Hz that is consistent with the known

1359: properties of HVC~neurons and of the HVC

1360: microcircuitry~\cite{Mooney05}. Although our results

1361: show that a brief burst can propagate in a stable

1362: manner, we and others~\cite{Jin05} have not been able

1363: to construct neuronal models, or simple networks of

1364: such models, that burst as rapidly and as briefly as

1365: actual \hvcra~neurons.

1366: 

1367: \vspace{.2in}

1368: 

1369: Note: Toward the completion of this paper, the authors

1370: received a preprint by Jin et al whose results overlap

1371: with the present paper (the preprint is related to a

1372: recent conference presentation~\cite{Jin05}). These

1373: authors also investigated whether brief high-frequency

1374: bursts analogous to those observed in \hvcra~neurons

1375: could propagate through a purely excitatory network,

1376: and they too found that this was possible for a range

1377: of parameters, thereby supporting the hypothesis that

1378: the bursting of \hvcra~neurons could be intrinsic

1379: to~HVC and causal.

1380: 

1381: The main difference of the present paper with the

1382: preprint is that we explored whether bursts could

1383: propagate under the simplest circumstances of a 1d,

1384: homogeneous, noiseless, purely-feedforward chain using

1385: simple neuronal models.  Jin et al studied a more

1386: complicated generalized (not purely feedforward),

1387: heterogeneous, noisy, purely excitatory synfire chain

1388: with many neurons per pool. Jin et al also used a more

1389: complicated two-compartment conductance-based neuronal

1390: model that included a low-threshold potassium channel

1391: that provided a strong spike-frequency adaptation

1392: similar to that observed experimentally. (This helped

1393: to create brief high-frequency bursts.) They also

1394: included calcium dynamics in the dendrite compartment

1395: that could lead to the firing of calcium spikes. Jin et

1396: al were able to show that adding the low-threshold

1397: potassium channel and calcium dynamics made the

1398: propagation of bursts through their generalized synfire

1399: chain more robust (one did not have to tune parameters

1400: so carefully to obtain a stable propagating

1401: burst). They also studied how jitter in the burst

1402: evolved along the length of the synfire chain, which we

1403: have also done in a forthcoming paper with a somewhat

1404: different model~\cite{Li06II}.

1405: 

1406: Given that our calculations and those of Jin et al have

1407: left out the role of inhibitory neurons and were based

1408: on a limited knowledge of~HVC neuronal properties and

1409: of the HVC microcircuitry (e.g., there is no evidence

1410: yet for a synfire chain in HVC, nor have the

1411: conductances of HVC~neurons been fully characterized),

1412: both calculations should be considered as modest but

1413: complementary and useful steps toward understanding

1414: possible mechanisms for the dynamics of~HVC during

1415: singing.

1416: 

1417: \appendix

1418: 

1419: \section{Hodgkin-Huxley Equations and Parameters}

1420: 

1421: \label{appendix:equations-parameters}

1422: 

1423: The single-compartment HH~model of

1424: Eq.~(\ref{eq:HH-model}) had the following five

1425: representative currents~$I_i(t)$:

1426: \begin{eqnarray}

1427:   I_{\rm leak} &=& \bar{g}_L (v-E_L), \nonumber \\

1428:   I_{\rm Na} &=& \bar{g}_{\rm Na} m^3 h (v-E_{\rm Na}), \nonumber\\

1429:   I_{\rm K} &=& \bar{g}_{\rm k} n^4 (v- E_{\rm K}),   \label{eq:hh-currents}

1430:  \\

1431:   I_{\rm Ca} &=& \bar{g}_{\rm Ca} m_1^2 h_1 (v-E_{\rm Ca}),\nonumber \\

1432:   I_{\rm KCa} &=& \bar{g}_{\rm KCa} n_1^3 (v-E_{\rm KCa}),\nonumber

1433: \end{eqnarray}

1434: with maximum conductances per unit area y (units of

1435: $\rm mS/mm^2$) given

1436: \begin{equation}

1437: \label{eq:hh-max-gbar-values}

1438:   \bar{g}_L = 0.003, \quad

1439:   \bar{g}_{\rm Na} = 1.0 , \quad

1440:   \bar{g}_{\rm K} = 0.22 , \qquad

1441:   \bar{g}_{\rm Ca} = 0.031 , \qquad

1442:   \bar{g}_{\rm KCa} = 0.007 , \qquad

1443: \end{equation}

1444: and corresponding reversal potentials (units of~mV):

1445: \begin{equation}

1446: \label{eq:hh-rev-potentials}

1447:   E_L = -70, \quad

1448:   E_{\rm Na} = 50 , \quad

1449:   E_{\rm K} = -80 , \qquad

1450:   E_{\rm Ca} = 120 , \qquad

1451:   E_{\rm KCa} = -80 . \qquad

1452: \end{equation}

1453: 

1454: \begin{table}[tbh]

1455:   \label{table:hh-parameters}

1456:   \begin{tabular}{c|c|c|c|c|c|c|}

1457:      \multicolumn{1}{c}{} 

1458:    & \multicolumn{1}{c}{\mbox{\ \ \ $m_{\ }$\ \ \ } }

1459:    & \multicolumn{1}{c}{\mbox{\ \ \ $n_{\ }$\ \ \  } }

1460:    & \multicolumn{1}{c}{\mbox{\ \ \  $h_{\ }$\ \ \  } }

1461:    & \multicolumn{1}{c}{\mbox{\ \ \  $m_1$\ \ \ }}

1462:    & \multicolumn{1}{c}{\mbox{\ \ \ $h_1$\ \ \  }}

1463:    & \multicolumn{1}{c}{\mbox{\ \ \ $n_1$\ \ \  }}\\

1464:     $v_0$ mV \, & -25.5 & -12.3  &  -60  & -50 & -80 &  \\

1465:     $v_1$ mV \, & 5.29 & 11.8 & -15 & 5 & -5 &  \\

1466:     $v_2$ mV \, & -120 & -28.3 & -62.9 & -50 & -80 & -35 \\

1467:     $v_3$ mV \, & 25 & 19.2 & 10 & 10 & 10 & 10 \\

1468:     $t_1$ ms \, & -1.26 & -6.4 & 1.1 & 20 & 60 & 30\\

1469:     $t_2$ ms \, & 1.32 & 7.2 & 0  & 10 & 10 & 10

1470:   \end{tabular}

1471:   \caption{Parameter values for the Hodgkin-Huxley model

1472:          Eq.~(\ref{eq:HH-model}) with currents

1473:   Eq.~(\ref{eq:hh-currents}) and gating equations

1474:   Eqs.~(\ref{eq:x-sigmoid-fn})-(\ref{eq:ca-tau-form}).

1475:   }

1476: \end{table}

1477: 

1478: The currents Eqs.~(\ref{eq:hh-currents}) depend on the

1479: membrane voltage~$v(t)$ and on gating variables $m$,

1480: $h$, $n$, $m_1$, $h_1$, and~$n_1$ that determine the

1481: probability as a function of time for certain channel

1482: subunits to be open. All the gating variables~$x(v)$

1483: except the KCa activation variable~$n_1$ obey the

1484: evolution equation

1485: \begin{equation}

1486:   \label{eq:gating-variable-eq}

1487:   \tau(v) \frac{dx}{dt} = x_{\infty}(v) - x  ,

1488: \end{equation}

1489: with the asymptotic value~$x_\infty(v)$ having the

1490: sigmoidal form

1491: \begin{equation}

1492:   x_{\infty}(v) = \left( 

1493:     1 + e^{\displaystyle -(v-v_0)/v_1} \right)^{-1} .

1494:   \label{eq:x-sigmoid-fn}

1495: \end{equation}

1496: The time constant~$\tau(v)$ has the form

1497: \begin{equation}

1498:   \tau(v) = t_2 + 

1499:   t_1 \left(

1500:     1 + e^{\displaystyle

1501:     -\left( (v-v_2)/v_3 \right) }

1502:    \right)^{-1} .

1503:    \label{eq:hh-tau-fn}

1504: \end{equation}

1505: for the standard HH channels~$m$, $h$, and~$n$, and the form

1506: \begin{equation}

1507:   \label{eq:ca-tau-form}

1508:   \tau(v) = t_2 + 

1509:     t_1 \exp\left[  

1510:       -\left( \frac{v-v_2}{v_3} \right)^2 

1511:     \right] ,

1512: \end{equation}

1513: for the Ca channels~$m_1$ and~$h_1$ and KCa channel $n_1$.

1514: Table~\ref{table:hh-parameters} gives the parameter

1515: values  that we used in these various expressions.

1516: 

1517: The KCa~activation variable~$n_1$ obeys the evolution

1518: equation

1519: \begin{equation}

1520:   \tau_{\rm KCa} \frac{dn_1}{dt} = f\left([{\rm

1521:   Ca}]_i\right)-n_1 ,

1522: \end{equation}

1523: where the expression~$[{\rm Ca}]_i$ denotes the average

1524: intracellular calcium concentration and the

1525: function~$f$ is given by

1526: \begin{equation}

1527:   f(\cai) = \frac{\cai}{(\cai+0.001)(1+\exp(-(v+40)/20))} .

1528: \end{equation}

1529: The intracellular calcium concentration accumulates

1530: when the low threshold calcium channel is open and

1531: calcium ions flow from extracellular solution into

1532: intracellular solution.  The phenomenological model

1533: for~$\cai$ was taken from Ref.~\cite{Turrigiano1995}

1534: \begin{equation}

1535:   \cai = ([{\rm Ca}]_{i0} - \cai - F*(I_{ca}))/\tau_{Ca} ,

1536: \end{equation}

1537: where the equilibrium Calcium concentration~$[{\rm

1538:   Ca}]_{i0}$ had the value~.00005~mM and the

1539: current-density factor~$F$ had the value 3~${\rm

1540:   mM-cm}^2/{\rm mA}$, and the time constant~$\tau_{Ca}$

1541: had the value~200 ms.  The KCa channel was chosen to

1542: imitate a low-threshold transient calcium current

1543: observed in HVC neurons by Kubota~\cite{Kubota98}.

1544: %%% MengRu: how was the KCa channel ``chosen to

1545: %%%  imitate'', i.e., what did you fit?

1546: 

1547: 

1548: \begin{acknowledgments}

1549:   The authors thank Richard Mooney, Jonathan Prather,

1550:   Melissa Coleman, Michale Fee, Stephen Shea, George

1551:   Augustine, and Paul Tiesinga for helpful discussions,

1552:   and Craig Henriquez and Duke University's Center for

1553:   Neural Engineering for providing support for one of

1554:   the authors (M.~Li).

1555: \end{acknowledgments}

1556: 

1557: 

1558: %\input{figure-captions}

1559: 

1560: \begin{thebibliography}{56}

1561: \expandafter\ifx\csname natexlab\endcsname\relax\def\natexlab#1{#1}\fi

1562: \expandafter\ifx\csname bibnamefont\endcsname\relax

1563:   \def\bibnamefont#1{#1}\fi

1564: \expandafter\ifx\csname bibfnamefont\endcsname\relax

1565:   \def\bibfnamefont#1{#1}\fi

1566: \expandafter\ifx\csname citenamefont\endcsname\relax

1567:   \def\citenamefont#1{#1}\fi

1568: \expandafter\ifx\csname url\endcsname\relax

1569:   \def\url#1{\texttt{#1}}\fi

1570: \expandafter\ifx\csname urlprefix\endcsname\relax\def\urlprefix{URL }\fi

1571: \providecommand{\bibinfo}[2]{#2}

1572: \providecommand{\eprint}[2][]{\url{#2}}

1573: 

1574: \bibitem[{\citenamefont{Doupe and Kuhl}(1999)}]{Doupe99}

1575: \bibinfo{author}{\bibfnamefont{A.~J.} \bibnamefont{Doupe}} \bibnamefont{and}

1576:   \bibinfo{author}{\bibfnamefont{P.~K.} \bibnamefont{Kuhl}},

1577:   \bibinfo{journal}{Annu. Rev. Neurosci.} \textbf{\bibinfo{volume}{22}},

1578:   \bibinfo{pages}{567} (\bibinfo{year}{1999}).

1579: 

1580: \bibitem[{\citenamefont{Catchpole and Slater}(1995)}]{Catchpole95}

1581: \bibinfo{author}{\bibfnamefont{C.~K.} \bibnamefont{Catchpole}}

1582:   \bibnamefont{and} \bibinfo{author}{\bibfnamefont{P.~J.~B.}

1583:   \bibnamefont{Slater}}, \emph{\bibinfo{title}{Bird Song: Biological themes and

1584:   variations}} (\bibinfo{publisher}{Cambridge University Press},

1585:   \bibinfo{address}{New York}, \bibinfo{year}{1995}).

1586: 

1587: \bibitem[{\citenamefont{Johnson et~al.}(2002)\citenamefont{Johnson, Soderstrom,

1588:   and Whitney}}]{Johnson02}

1589: \bibinfo{author}{\bibfnamefont{F.}~\bibnamefont{Johnson}},

1590:   \bibinfo{author}{\bibfnamefont{K.}~\bibnamefont{Soderstrom}},

1591:   \bibnamefont{and} \bibinfo{author}{\bibfnamefont{O.}~\bibnamefont{Whitney}},

1592:   \bibinfo{journal}{Behavioural Brain Research} \textbf{\bibinfo{volume}{131}},

1593:   \bibinfo{pages}{57} (\bibinfo{year}{2002}).

1594: 

1595: \bibitem[{\citenamefont{Hahnloser et~al.}(2002)\citenamefont{Hahnloser,

1596:   Kozhevnikov, and Fee}}]{Hahnloser02}

1597: \bibinfo{author}{\bibfnamefont{R.~H.~R.} \bibnamefont{Hahnloser}},

1598:   \bibinfo{author}{\bibfnamefont{A.~A.} \bibnamefont{Kozhevnikov}},

1599:   \bibnamefont{and} \bibinfo{author}{\bibfnamefont{M.~S.} \bibnamefont{Fee}},

1600:   \bibinfo{journal}{Nature} \textbf{\bibinfo{volume}{419}}, \bibinfo{pages}{65}

1601:   (\bibinfo{year}{2002}).

1602: 

1603: \bibitem[{\citenamefont{Reiner et~al.}(2004)\citenamefont{Reiner, Perkel,

1604:   Mello, and Jarvis}}]{Reiner04}

1605: \bibinfo{author}{\bibfnamefont{A.}~\bibnamefont{Reiner}},

1606:   \bibinfo{author}{\bibfnamefont{D.~J.} \bibnamefont{Perkel}},

1607:   \bibinfo{author}{\bibfnamefont{C.~V.} \bibnamefont{Mello}}, \bibnamefont{and}

1608:   \bibinfo{author}{\bibfnamefont{E.~D.} \bibnamefont{Jarvis}},

1609:   \bibinfo{journal}{Ann. N. Y. Acad. Sci.} \textbf{\bibinfo{volume}{1016}},

1610:   \bibinfo{pages}{77} (\bibinfo{year}{2004}).

1611: 

1612: \bibitem[{\citenamefont{Fee et~al.}(2004)\citenamefont{Fee, Kozhevnikov, and

1613:   Hahnloser}}]{Fee04}

1614: \bibinfo{author}{\bibfnamefont{M.~S.} \bibnamefont{Fee}},

1615:   \bibinfo{author}{\bibfnamefont{A.}~\bibnamefont{Kozhevnikov}},

1616:   \bibnamefont{and} \bibinfo{author}{\bibfnamefont{R.~H.~R.}

1617:   \bibnamefont{Hahnloser}}, \bibinfo{journal}{Ann. N. Y. Acad. Sci.}

1618:   \textbf{\bibinfo{volume}{1016}}, \bibinfo{pages}{153} (\bibinfo{year}{2004}).

1619: 

1620: \bibitem[{\citenamefont{Jin et~al.}(2005)\citenamefont{Jin, Ramazanoglu, and

1621:   Seung}}]{Jin05}

1622: \bibinfo{author}{\bibfnamefont{D.}~\bibnamefont{Jin}},

1623:   \bibinfo{author}{\bibfnamefont{F.}~\bibnamefont{Ramazanoglu}},

1624:   \bibnamefont{and} \bibinfo{author}{\bibfnamefont{H.~S.} \bibnamefont{Seung}}

1625:   (\bibinfo{year}{2005}), \bibinfo{note}{society for Neuroscience Abstracts

1626:   No.~753.15}.

1627: 

1628: \bibitem[{\citenamefont{Abeles}(1991)}]{Abeles91}

1629: \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Abeles}},

1630:   \emph{\bibinfo{title}{Corticonics: Neural Circuits of the Cerebral Cortex}}

1631:   (\bibinfo{publisher}{Cambridge University Press},

1632:   \bibinfo{address}{Cambridge, England}, \bibinfo{year}{1991}).

1633: 

1634: \bibitem[{\citenamefont{Diesmann et~al.}(1999)\citenamefont{Diesmann, Gewaltig,

1635:   and Aertsen}}]{Diesmann99}

1636: \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Diesmann}},

1637:   \bibinfo{author}{\bibfnamefont{M.-O.} \bibnamefont{Gewaltig}},

1638:   \bibnamefont{and} \bibinfo{author}{\bibfnamefont{A.}~\bibnamefont{Aertsen}},

1639:   \bibinfo{journal}{Nature} \textbf{\bibinfo{volume}{402}},

1640:   \bibinfo{pages}{529} (\bibinfo{year}{1999}).

1641: 

1642: \bibitem[{\citenamefont{O\c{s}an et~al.}(2004)\citenamefont{O\c{s}an, Curtu,

1643:   Rubin, and Ermentrout}}]{Osan04}

1644: \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{O\c{s}an}},

1645:   \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Curtu}},

1646:   \bibinfo{author}{\bibfnamefont{J.}~\bibnamefont{Rubin}}, \bibnamefont{and}

1647:   \bibinfo{author}{\bibfnamefont{B.}~\bibnamefont{Ermentrout}},

1648:   \bibinfo{journal}{J. Math. Biol.} \textbf{\bibinfo{volume}{48}},

1649:   \bibinfo{pages}{243} (\bibinfo{year}{2004}).

1650: 

1651: \bibitem[{\citenamefont{Reyes}(2003)}]{Reyes03}

1652: \bibinfo{author}{\bibfnamefont{A.~D.} \bibnamefont{Reyes}},

1653:   \bibinfo{journal}{Nature Neuroscience} \textbf{\bibinfo{volume}{6}},

1654:   \bibinfo{pages}{593} (\bibinfo{year}{2003}).

1655: 

1656: \bibitem[{\citenamefont{Golomb and Amitai}(1997)}]{Golomb97}

1657: \bibinfo{author}{\bibfnamefont{D.}~\bibnamefont{Golomb}} \bibnamefont{and}

1658:   \bibinfo{author}{\bibfnamefont{Y.}~\bibnamefont{Amitai}},

1659:   \bibinfo{journal}{J. Neurophysiol.} \textbf{\bibinfo{volume}{78}},

1660:   \bibinfo{pages}{1199} (\bibinfo{year}{1997}).

1661: 

1662: \bibitem[{\citenamefont{Litvak et~al.}(2003)\citenamefont{Litvak, Sompolinsky,

1663:   Segev, and Abeles}}]{Litvak03}

1664: \bibinfo{author}{\bibfnamefont{V.}~\bibnamefont{Litvak}},

1665:   \bibinfo{author}{\bibfnamefont{H.}~\bibnamefont{Sompolinsky}},

1666:   \bibinfo{author}{\bibfnamefont{I.}~\bibnamefont{Segev}}, \bibnamefont{and}

1667:   \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Abeles}},

1668:   \bibinfo{journal}{Journal of Neuroscience} \textbf{\bibinfo{volume}{23}},

1669:   \bibinfo{pages}{3096} (\bibinfo{year}{2003}).

1670: 

1671: \bibitem[{\citenamefont{Vogels and Abbott}(2095)}]{Vogels05}

1672: \bibinfo{author}{\bibfnamefont{T.~P.} \bibnamefont{Vogels}} \bibnamefont{and}

1673:   \bibinfo{author}{\bibfnamefont{L.~E.} \bibnamefont{Abbott}},

1674:   \bibinfo{journal}{J. of Neuroscience} \textbf{\bibinfo{volume}{25}},

1675:   \bibinfo{pages}{10786} (\bibinfo{year}{2095}).

1676: 

1677: \bibitem[{\citenamefont{Mooney}(2000)}]{Mooney00}

1678: \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Mooney}}, \bibinfo{journal}{J.

1679:   Neurosci.} \textbf{\bibinfo{volume}{20}}, \bibinfo{pages}{5420}

1680:   (\bibinfo{year}{2000}).

1681: 

1682: \bibitem[{\citenamefont{Ohki et~al.}(2005)\citenamefont{Ohki, Chung, Ch'ng,

1683:   Kara, and Reid}}]{Ohki05}

1684: \bibinfo{author}{\bibfnamefont{K.}~\bibnamefont{Ohki}},

1685:   \bibinfo{author}{\bibfnamefont{S.}~\bibnamefont{Chung}},

1686:   \bibinfo{author}{\bibfnamefont{Y.~H.} \bibnamefont{Ch'ng}},

1687:   \bibinfo{author}{\bibfnamefont{P.}~\bibnamefont{Kara}}, \bibnamefont{and}

1688:   \bibinfo{author}{\bibfnamefont{R.~C.} \bibnamefont{Reid}},

1689:   \bibinfo{journal}{Nature} \textbf{\bibinfo{volume}{433}},

1690:   \bibinfo{pages}{597} (\bibinfo{year}{2005}).

1691: 

1692: \bibitem[{\citenamefont{Li and Greenside}(2006)}]{Li06II}

1693: \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Li}} \bibnamefont{and}

1694:   \bibinfo{author}{\bibfnamefont{H.}~\bibnamefont{Greenside}}

1695:   (\bibinfo{year}{2006}), \bibinfo{note}{in preparation.}

1696: 

1697: \bibitem[{\citenamefont{Li}(2006)}]{Li06thesis}

1698: \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Li}}, Ph.D. thesis,

1699:   \bibinfo{school}{Duke University} (\bibinfo{year}{2006}).

1700: 

1701: \bibitem[{\citenamefont{Kubota and Taniguchi}(1998)}]{Kubota98}

1702: \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Kubota}} \bibnamefont{and}

1703:   \bibinfo{author}{\bibfnamefont{I.}~\bibnamefont{Taniguchi}},

1704:   \bibinfo{journal}{J. Neurophysiol.} \textbf{\bibinfo{volume}{80}},

1705:   \bibinfo{pages}{914} (\bibinfo{year}{1998}).

1706: 

1707: \bibitem[{\citenamefont{Mooney and Prather}(2005)}]{Mooney05}

1708: \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Mooney}} \bibnamefont{and}

1709:   \bibinfo{author}{\bibfnamefont{J.~F.} \bibnamefont{Prather}},

1710:   \bibinfo{journal}{Journal of Neuroscience} \textbf{\bibinfo{volume}{25}},

1711:   \bibinfo{pages}{1952} (\bibinfo{year}{2005}).

1712: 

1713: \bibitem[{\citenamefont{Dutar et~al.}(1998)\citenamefont{Dutar, Vu, and

1714:   Perkel}}]{Dutar98}

1715: \bibinfo{author}{\bibfnamefont{P.}~\bibnamefont{Dutar}},

1716:   \bibinfo{author}{\bibfnamefont{H.~M.} \bibnamefont{Vu}}, \bibnamefont{and}

1717:   \bibinfo{author}{\bibfnamefont{D.~J.} \bibnamefont{Perkel}},

1718:   \bibinfo{journal}{J. Neurophysiol.} \textbf{\bibinfo{volume}{80}},

1719:   \bibinfo{pages}{1828} (\bibinfo{year}{1998}).

1720: 

1721: \bibitem[{\citenamefont{Wild et~al.}(2005)\citenamefont{Wild, Williams, Howie,

1722:   and Mooney}}]{Wild05}

1723: \bibinfo{author}{\bibfnamefont{J.~M.} \bibnamefont{Wild}},

1724:   \bibinfo{author}{\bibfnamefont{M.~N.} \bibnamefont{Williams}},

1725:   \bibinfo{author}{\bibfnamefont{G.~J.} \bibnamefont{Howie}}, \bibnamefont{and}

1726:   \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Mooney}},

1727:   \bibinfo{journal}{Journal of Comparative Neurology}

1728:   \textbf{\bibinfo{volume}{483}}, \bibinfo{pages}{76} (\bibinfo{year}{2005}).

1729: 

1730: \bibitem[{\citenamefont{Traub et~al.}(1999)\citenamefont{Traub, Jefferys, and

1731:   Whittington}}]{Traub99}

1732: \bibinfo{author}{\bibfnamefont{R.~D.} \bibnamefont{Traub}},

1733:   \bibinfo{author}{\bibfnamefont{J.~G.~R.} \bibnamefont{Jefferys}},

1734:   \bibnamefont{and} \bibinfo{author}{\bibfnamefont{M.~A.}

1735:   \bibnamefont{Whittington}}, \emph{\bibinfo{title}{Fast Oscillations in

1736:   Cortical Circuits}} (\bibinfo{publisher}{Bradford Books},

1737:   \bibinfo{address}{New York}, \bibinfo{year}{1999}).

1738: 

1739: \bibitem[{\citenamefont{Nottebohm}(2002)}]{NottebohmJN02}

1740: \bibinfo{author}{\bibfnamefont{F.}~\bibnamefont{Nottebohm}},

1741:   \bibinfo{journal}{J Neuroscience} \textbf{\bibinfo{volume}{22}},

1742:   \bibinfo{pages}{624} (\bibinfo{year}{2002}).

1743: 

1744: \bibitem[{\citenamefont{Wang et~al.}(2002)\citenamefont{Wang, Hurley, Pytte,

1745:   and Kirn}}]{Wang02}

1746: \bibinfo{author}{\bibfnamefont{N.}~\bibnamefont{Wang}},

1747:   \bibinfo{author}{\bibfnamefont{P.}~\bibnamefont{Hurley}},

1748:   \bibinfo{author}{\bibfnamefont{T.}~\bibnamefont{Pytte}}, \bibnamefont{and}

1749:   \bibinfo{author}{\bibfnamefont{J.~R.} \bibnamefont{Kirn}},

1750:   \bibinfo{journal}{J. Neurosci.} \textbf{\bibinfo{volume}{22}},

1751:   \bibinfo{pages}{10864} (\bibinfo{year}{2002}).

1752: 

1753: \bibitem[{\citenamefont{Ikegaya et~al.}(2004)\citenamefont{Ikegaya, Aaron,

1754:   Cossart, Aronov, Lampl, Ferster, and Yuste}}]{Ikegaya04}

1755: \bibinfo{author}{\bibfnamefont{Y.}~\bibnamefont{Ikegaya}},

1756:   \bibinfo{author}{\bibfnamefont{G.}~\bibnamefont{Aaron}},

1757:   \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Cossart}},

1758:   \bibinfo{author}{\bibfnamefont{D.}~\bibnamefont{Aronov}},

1759:   \bibinfo{author}{\bibfnamefont{I.}~\bibnamefont{Lampl}},

1760:   \bibinfo{author}{\bibfnamefont{D.}~\bibnamefont{Ferster}}, \bibnamefont{and}

1761:   \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Yuste}},

1762:   \bibinfo{journal}{Science} \textbf{\bibinfo{volume}{304}},

1763:   \bibinfo{pages}{559} (\bibinfo{year}{2004}).

1764: 

1765: \bibitem[{\citenamefont{Leonardo and Fee}(2005)}]{Leonardo05}

1766: \bibinfo{author}{\bibfnamefont{A.}~\bibnamefont{Leonardo}} \bibnamefont{and}

1767:   \bibinfo{author}{\bibfnamefont{M.~S.} \bibnamefont{Fee}},

1768:   \bibinfo{journal}{Journal of Neuroscience} \textbf{\bibinfo{volume}{25}},

1769:   \bibinfo{pages}{652} (\bibinfo{year}{2005}).

1770: 

1771: \bibitem[{\citenamefont{Doya and Sejnowski}(1998)}]{Doya98}

1772: \bibinfo{author}{\bibfnamefont{K.}~\bibnamefont{Doya}} \bibnamefont{and}

1773:   \bibinfo{author}{\bibfnamefont{T.~J.} \bibnamefont{Sejnowski}}, in

1774:   \emph{\bibinfo{booktitle}{Central Auditory Processing and Neural Modeling}},

1775:   edited by \bibinfo{editor}{\bibnamefont{Poon}} \bibnamefont{and}

1776:   \bibinfo{editor}{\bibnamefont{Brugge}} (\bibinfo{publisher}{Plenum Press},

1777:   \bibinfo{address}{New York}, \bibinfo{year}{1998}).

1778: 

1779: \bibitem[{\citenamefont{Abarbanel et~al.}(2004)\citenamefont{Abarbanel, Gibb,

1780:   Mindlin, and Talathi}}]{Abarbanel04jn}

1781: \bibinfo{author}{\bibfnamefont{H.~D.~I.} \bibnamefont{Abarbanel}},

1782:   \bibinfo{author}{\bibfnamefont{L.}~\bibnamefont{Gibb}},

1783:   \bibinfo{author}{\bibfnamefont{G.~B.} \bibnamefont{Mindlin}},

1784:   \bibnamefont{and} \bibinfo{author}{\bibfnamefont{S.}~\bibnamefont{Talathi}},

1785:   \bibinfo{journal}{J. Neurophysiol.} \textbf{\bibinfo{volume}{92}},

1786:   \bibinfo{pages}{96} (\bibinfo{year}{2004}).

1787: 

1788: \bibitem[{\citenamefont{Berger}(2004)}]{Berger04}

1789: \bibinfo{author}{\bibfnamefont{L.}~\bibnamefont{Berger}}, Master's thesis,

1790:   \bibinfo{school}{Institute of Neuroinformatics, University/ETH Z\"urich}

1791:   (\bibinfo{year}{2004}).

1792: 

1793: \bibitem[{\citenamefont{Drew and Abbott}(2003)}]{Drew03}

1794: \bibinfo{author}{\bibfnamefont{P.~J.} \bibnamefont{Drew}} \bibnamefont{and}

1795:   \bibinfo{author}{\bibfnamefont{L.~F.} \bibnamefont{Abbott}},

1796:   \bibinfo{journal}{J. Neurophysiol.} \textbf{\bibinfo{volume}{89}},

1797:   \bibinfo{pages}{2697} (\bibinfo{year}{2003}).

1798: 

1799: \bibitem[{\citenamefont{Kleinfeld and Sompolinsky}(1989)}]{Kleinfeld89}

1800: \bibinfo{author}{\bibfnamefont{D.}~\bibnamefont{Kleinfeld}} \bibnamefont{and}

1801:   \bibinfo{author}{\bibfnamefont{H.}~\bibnamefont{Sompolinsky}}, in

1802:   \emph{\bibinfo{booktitle}{In Methods in Neuronal Modeling: From Synapse to

1803:   Networks}}, edited by \bibinfo{editor}{\bibfnamefont{C.}~\bibnamefont{Koch}}

1804:   \bibnamefont{and} \bibinfo{editor}{\bibfnamefont{I.}~\bibnamefont{Segev}}

1805:   (\bibinfo{publisher}{MIT Press}, \bibinfo{address}{Cambridge, MA},

1806:   \bibinfo{year}{1989}), pp. \bibinfo{pages}{195--246}.

1807: 

1808: \bibitem[{\citenamefont{Huerta and Rabinovich}(2004)}]{Huerta04}

1809: \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Huerta}} \bibnamefont{and}

1810:   \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Rabinovich}},

1811:   \bibinfo{journal}{Phys. Rev. Lett.} \textbf{\bibinfo{volume}{93}},

1812:   \bibinfo{pages}{238104 (4 pages)} (\bibinfo{year}{2004}).

1813: 

1814: \bibitem[{\citenamefont{Dayan and Abbott}(2001)}]{Dayan01}

1815: \bibinfo{author}{\bibfnamefont{P.}~\bibnamefont{Dayan}} \bibnamefont{and}

1816:   \bibinfo{author}{\bibfnamefont{L.~F.} \bibnamefont{Abbott}},

1817:   \emph{\bibinfo{title}{Theoretical Neuroscience: Computational and

1818:   Mathematical Modeling of Neural Systems}} (\bibinfo{publisher}{MIT Press},

1819:   \bibinfo{year}{2001}).

1820: 

1821: \bibitem[{\citenamefont{Kincaid and Cheney}(1996)}]{Kincaid96}

1822: \bibinfo{author}{\bibfnamefont{D.}~\bibnamefont{Kincaid}} \bibnamefont{and}

1823:   \bibinfo{author}{\bibfnamefont{W.}~\bibnamefont{Cheney}},

1824:   \emph{\bibinfo{title}{Numerical Analysis}} (\bibinfo{publisher}{Brooks/Cole},

1825:   \bibinfo{address}{Pacific Grove, California}, \bibinfo{year}{1996}),

1826:   \bibinfo{edition}{2nd} ed.

1827: 

1828: \bibitem[{\citenamefont{MathWorks}()}]{Matlab}

1829: \bibinfo{author}{\bibnamefont{MathWorks}}, \emph{\bibinfo{title}{Matlab}},

1830:   \bibinfo{howpublished}{http://www.mathworks.com/products/matlab/}.

1831: 

1832: \bibitem[{\citenamefont{Kubota and Saito}(1991)}]{Kubota91}

1833: \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Kubota}} \bibnamefont{and}

1834:   \bibinfo{author}{\bibfnamefont{N.}~\bibnamefont{Saito}},

1835:   \bibinfo{journal}{Journal of Physiology} \textbf{\bibinfo{volume}{440}},

1836:   \bibinfo{pages}{131} (\bibinfo{year}{1991}).

1837: 

1838: \bibitem[{\citenamefont{Prinz et~al.}(2003)\citenamefont{Prinz, Billimoria, and

1839:   Marder}}]{Prinz03}

1840: \bibinfo{author}{\bibfnamefont{A.~A.} \bibnamefont{Prinz}},

1841:   \bibinfo{author}{\bibfnamefont{C.~P.} \bibnamefont{Billimoria}},

1842:   \bibnamefont{and} \bibinfo{author}{\bibfnamefont{E.}~\bibnamefont{Marder}},

1843:   \bibinfo{journal}{J. Neurophysiol.} \textbf{\bibinfo{volume}{90}},

1844:   \bibinfo{pages}{3998} (\bibinfo{year}{2003}).

1845: 

1846: \bibitem[{\citenamefont{Huguenard and McCormick}(1992)}]{Huguenard92}

1847: \bibinfo{author}{\bibfnamefont{J.}~\bibnamefont{Huguenard}} \bibnamefont{and}

1848:   \bibinfo{author}{\bibfnamefont{D.}~\bibnamefont{McCormick}},

1849:   \bibinfo{journal}{J. Neurophysiology} \textbf{\bibinfo{volume}{68}},

1850:   \bibinfo{pages}{1373} (\bibinfo{year}{1992}).

1851: 

1852: \bibitem[{\citenamefont{Koch and Segev}(1998)}]{Yamada98}

1853: \bibinfo{editor}{\bibfnamefont{C.}~\bibnamefont{Koch}} \bibnamefont{and}

1854:   \bibinfo{editor}{\bibfnamefont{I.}~\bibnamefont{Segev}}, eds.,

1855:   \emph{\bibinfo{title}{Multiple Channels and Calcium Dynamics}}

1856:   (\bibinfo{publisher}{MIT}, \bibinfo{year}{1998}).

1857: 

1858: \bibitem[{\citenamefont{Hines et~al.}()\citenamefont{Hines, Moore, and

1859:   Carnevale}}]{NeuronCode}

1860: \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Hines}},

1861:   \bibinfo{author}{\bibfnamefont{J.~W.} \bibnamefont{Moore}}, \bibnamefont{and}

1862:   \bibinfo{author}{\bibfnamefont{T.}~\bibnamefont{Carnevale}},

1863:   \emph{\bibinfo{title}{Neuron}},

1864:   \bibinfo{howpublished}{http://neuron.duke.edu/}.

1865: 

1866: \bibitem[{\citenamefont{Hertz et~al.}(1991)\citenamefont{Hertz, Krogh, and

1867:   Palmer}}]{Hertz91}

1868: \bibinfo{author}{\bibfnamefont{J.}~\bibnamefont{Hertz}},

1869:   \bibinfo{author}{\bibfnamefont{A.}~\bibnamefont{Krogh}}, \bibnamefont{and}

1870:   \bibinfo{author}{\bibfnamefont{R.~G.} \bibnamefont{Palmer}},

1871:   \emph{\bibinfo{title}{Introduction to the Theory of Neural Computation}},

1872:   no.~\bibinfo{number}{I} in \bibinfo{series}{Santa Fe Institute Studies in the

1873:   Sciences of Complexity} (\bibinfo{publisher}{Westview},

1874:   \bibinfo{address}{Boulder, CO}, \bibinfo{year}{1991}).

1875: 

1876: \bibitem[{\citenamefont{Hopfield}(1982)}]{Hopfield82}

1877: \bibinfo{author}{\bibfnamefont{J.~J.} \bibnamefont{Hopfield}},

1878:   \bibinfo{journal}{Proc. Nat. Acad. Sci.} \textbf{\bibinfo{volume}{79}},

1879:   \bibinfo{pages}{2554} (\bibinfo{year}{1982}).

1880: 

1881: \bibitem[{\citenamefont{Coleman and Mooney}(2004)}]{Coleman04}

1882: \bibinfo{author}{\bibfnamefont{M.~J.} \bibnamefont{Coleman}} \bibnamefont{and}

1883:   \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Mooney}},

1884:   \bibinfo{journal}{Journal of Neuroscience} \textbf{\bibinfo{volume}{24}},

1885:   \bibinfo{pages}{7251} (\bibinfo{year}{2004}).

1886: 

1887: \bibitem[{\citenamefont{Rosen and Mooney}(2006)}]{Rosen06}

1888: \bibinfo{author}{\bibfnamefont{M.~J.} \bibnamefont{Rosen}} \bibnamefont{and}

1889:   \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Mooney}},

1890:   \bibinfo{journal}{J. Neurophysiol.} \textbf{\bibinfo{volume}{95}},

1891:   \bibinfo{pages}{1158} (\bibinfo{year}{2006}).

1892: 

1893: \bibitem[{\citenamefont{Solis and Perkel}(2005)}]{Solis05}

1894: \bibinfo{author}{\bibfnamefont{M.~M.} \bibnamefont{Solis}} \bibnamefont{and}

1895:   \bibinfo{author}{\bibfnamefont{D.~J.} \bibnamefont{Perkel}},

1896:   \bibinfo{journal}{J. Neuroscience} \textbf{\bibinfo{volume}{25}},

1897:   \bibinfo{pages}{2811} (\bibinfo{year}{2005}).

1898: 

1899: \bibitem[{\citenamefont{Fiete et~al.}(2004)\citenamefont{Fiete, Hahnloser, Fee,

1900:   and Sueng}}]{Fiete04}

1901: \bibinfo{author}{\bibfnamefont{I.~R.} \bibnamefont{Fiete}},

1902:   \bibinfo{author}{\bibfnamefont{R.~H.~R.} \bibnamefont{Hahnloser}},

1903:   \bibinfo{author}{\bibfnamefont{M.~S.} \bibnamefont{Fee}}, \bibnamefont{and}

1904:   \bibinfo{author}{\bibfnamefont{H.~S.} \bibnamefont{Sueng}},

1905:   \bibinfo{journal}{J. Neurophysiol.} \textbf{\bibinfo{volume}{92}},

1906:   \bibinfo{pages}{2274} (\bibinfo{year}{2004}).

1907: 

1908: \bibitem[{\citenamefont{Lisman}(1997)}]{Lisman97}

1909: \bibinfo{author}{\bibfnamefont{J.~E.} \bibnamefont{Lisman}},

1910:   \bibinfo{journal}{Trends Neurosci.} \textbf{\bibinfo{volume}{20}},

1911:   \bibinfo{pages}{38} (\bibinfo{year}{1997}).

1912: 

1913: \bibitem[{\citenamefont{de~la Prida et~al.}(1997)\citenamefont{de~la Prida,

1914:   Stollenwerk, and Sanchez-Andres}}]{Prida97}

1915: \bibinfo{author}{\bibfnamefont{L.~M.} \bibnamefont{de~la Prida}},

1916:   \bibinfo{author}{\bibfnamefont{N.}~\bibnamefont{Stollenwerk}},

1917:   \bibnamefont{and} \bibinfo{author}{\bibfnamefont{J.~V.}

1918:   \bibnamefont{Sanchez-Andres}}, \bibinfo{journal}{Physica D}

1919:   \textbf{\bibinfo{volume}{110}}, \bibinfo{pages}{323} (\bibinfo{year}{1997}).

1920: 

1921: \bibitem[{\citenamefont{Vu et~al.}(1994)\citenamefont{Vu, Mazurek, and

1922:   Kuo}}]{Vu94}

1923: \bibinfo{author}{\bibfnamefont{E.~T.} \bibnamefont{Vu}},

1924:   \bibinfo{author}{\bibfnamefont{M.~E.} \bibnamefont{Mazurek}},

1925:   \bibnamefont{and} \bibinfo{author}{\bibfnamefont{Y.-C.} \bibnamefont{Kuo}},

1926:   \bibinfo{journal}{J. Neuroscience} \textbf{\bibinfo{volume}{14}},

1927:   \bibinfo{pages}{6924} (\bibinfo{year}{1994}).

1928: 

1929: \bibitem[{\citenamefont{Vu et~al.}(1998)\citenamefont{Vu, Schmidt, and

1930:   Mazurek}}]{Vu98}

1931: \bibinfo{author}{\bibfnamefont{E.~T.} \bibnamefont{Vu}},

1932:   \bibinfo{author}{\bibfnamefont{M.~F.} \bibnamefont{Schmidt}},

1933:   \bibnamefont{and} \bibinfo{author}{\bibfnamefont{M.~E.}

1934:   \bibnamefont{Mazurek}}, \bibinfo{journal}{J. Neuroscience}

1935:   \textbf{\bibinfo{volume}{18}}, \bibinfo{pages}{9088} (\bibinfo{year}{1998}).

1936: 

1937: \bibitem[{\citenamefont{Vicario and Simpson}(1995)}]{Vicario95}

1938: \bibinfo{author}{\bibfnamefont{D.~S.} \bibnamefont{Vicario}} \bibnamefont{and}

1939:   \bibinfo{author}{\bibfnamefont{H.~B.} \bibnamefont{Simpson}},

1940:   \bibinfo{journal}{J. Neurophys.} \textbf{\bibinfo{volume}{73}},

1941:   \bibinfo{pages}{2602} (\bibinfo{year}{1995}).

1942: 

1943: \bibitem[{\citenamefont{Losonczy and Magee}(2006)}]{Losonczy06}

1944: \bibinfo{author}{\bibfnamefont{A.}~\bibnamefont{Losonczy}} \bibnamefont{and}

1945:   \bibinfo{author}{\bibfnamefont{J.~C.} \bibnamefont{Magee}},

1946:   \bibinfo{journal}{Neuron} \textbf{\bibinfo{volume}{50}}, \bibinfo{pages}{291}

1947:   (\bibinfo{year}{2006}).

1948: 

1949: \bibitem[{\citenamefont{Berg\'e et~al.}(1984)\citenamefont{Berg\'e, Pomeau, and

1950:   Vidal}}]{Berge84}

1951: \bibinfo{author}{\bibfnamefont{P.}~\bibnamefont{Berg\'e}},

1952:   \bibinfo{author}{\bibfnamefont{Y.}~\bibnamefont{Pomeau}}, \bibnamefont{and}

1953:   \bibinfo{author}{\bibfnamefont{C.}~\bibnamefont{Vidal}},

1954:   \emph{\bibinfo{title}{Order Within Chaos}} (\bibinfo{publisher}{John Wiley \&

1955:   Sons}, \bibinfo{address}{New York}, \bibinfo{year}{1984}).

1956: 

1957: \bibitem[{\citenamefont{Kopell and Ermentrout}(2004)}]{Kopell04}

1958: \bibinfo{author}{\bibfnamefont{N.}~\bibnamefont{Kopell}} \bibnamefont{and}

1959:   \bibinfo{author}{\bibfnamefont{B.}~\bibnamefont{Ermentrout}},

1960:   \bibinfo{journal}{PNAS} \textbf{\bibinfo{volume}{101}},

1961:   \bibinfo{pages}{15482} (\bibinfo{year}{2004}).

1962: 

1963: \bibitem[{\citenamefont{Turrigiano et~al.}(1995)\citenamefont{Turrigiano,

1964:   Lemasson, and Marder}}]{Turrigiano1995}

1965: \bibinfo{author}{\bibfnamefont{G.}~\bibnamefont{Turrigiano}},

1966:   \bibinfo{author}{\bibfnamefont{G.}~\bibnamefont{Lemasson}}, \bibnamefont{and}

1967:   \bibinfo{author}{\bibfnamefont{E.}~\bibnamefont{Marder}}, \bibinfo{journal}{J

1968:   Neuroscience} \textbf{\bibinfo{volume}{15}}, \bibinfo{pages}{3640}

1969:   (\bibinfo{year}{1995}).

1970: 

1971: \end{thebibliography}

1972: 

1973: \end{document}