Matthijs A. A. van der Meer, James J. Knierim, D. Yoganarasimha, Emma R.

Wood and Mark C. W. van Rossum

J Neurophysiol 98:1883-1897, 2007. First published Jun 27, 2007; doi:10.1152/jn.00233.2007

You might find this additional information useful...

This article cites 61 articles, 28 of which you can access free at:
[Link]

Updated information and services including high-resolution figures, can be found at:
[Link]

Additional material and information about Journal of Neurophysiology can be found at:
[Link]

This information is current as of May 9, 2008 .

Downloaded from [Link] on May 9, 2008

Journal of Neurophysiology publishes original articles on the function of the nervous system. It is published 12 times a year
(monthly) by the American Physiological Society, 9650 Rockville Pike, Bethesda MD 20814-3991. Copyright © 2005 by the
American Physiological Society. ISSN: 0022-3077, ESSN: 1522-1598. Visit our website at [Link]
 J Neurophysiol 98: 1883–1897, 2007.
First published June 27, 2007; doi:10.1152/jn.00233.2007.

Anticipation in the Rodent Head Direction System Can Be Explained by an

Interaction of Head Movements and Vestibular Firing Properties

Matthijs A. A. van der Meer,1 James J. Knierim,4 D. Yoganarasimha,4 Emma R. Wood,2

and Mark C. W. van Rossum3
1
Neuroinformatics Doctoral Training Centre, 2Centre for Cognitive and Neural Systems, and 3Institute for Adaptive and Neural
Computation, University of Edinburgh, Edinburgh, United Kingdom; and 4Department of Neurobiology and Anatomy, W. M. Keck Center
for the Neurobiology of Learning and Memory, The University of Texas Medical School, Houston, Texas
Submitted 4 March 2007; accepted in final form 25 June 2007

van der Meer MA, Knierim JJ, Yoganarasimha D, Wood ER, van HD system is thought to be a critical component for several
Rossum MC. Anticipation in the rodent head direction system can be forms of spatial navigation (Gallistel 1990; McNaughton et al.
explained by an interaction of head movements and vestibular firing 1991, 1996, 2006; Redish 1999; Taube 1998; but see Dud-
properties. J Neurophysiol 98: 1883–1897, 2007. First published June chenko et al. 2005; Muir and Taube 2002).
27, 2007; doi:10.1152/jn.00233.2007. The rodent head-direction (HD)
system, which codes for the animal’s head direction in the horizontal
When viewed as a sensory system, HD cells would be
expected to encode with some delay or time lag caused by

Downloaded from [Link] on May 9, 2008

plane, is thought to be critically involved in spatial navigation.
Electrophysiological recording studies have shown that HD cells can transduction, transmission, and other delays. However, HD cell
anticipate the animal’s HD by up to 75– 80 ms. The origin of this activity (in some brain areas) correlates best with future HD, as
anticipation is poorly understood. In this modeling study, we provide revealed by time-slide analyses (Blair and Sharp 1995; Taube
a novel explanation for HD anticipation that relies on the firing and Muller 1998). Following the HD literature, we refer to this
properties of neurons afferent to the HD system. By incorporating effect as anticipation; however, it should be noted that this does
spike rate adaptation and postinhibitory rebound as observed in not necessarily imply an active prediction process. Experimen-
medial vestibular nucleus neurons, our model produces realistic an- tally, anticipatory time intervals (ATIs) of up to 75– 80 ms
ticipation on a large corpus of rat movement data. In addition, HD have been observed (Blair et al. 1997; Stackman and Taube
anticipation varies between recording sessions of the same cell,
between active and passive movement, and between different studies.
1998; note erratum). These ATIs vary both between different
Such differences do not appear to be correlated with behavioral variables HD cells and between different recording sessions of the same
and cannot be accounted for using earlier models. In the present model, cell. Blair et al. (1997) report that ATIs of the same HD cell
anticipation depends on the power spectrum of the head movements. By recorded during multiple recording sessions can differ by as
direct comparison with recording data, we show that the model explains much as 50 ms, although on average, within-cell ATI variabil-
60 – 80% of the observed anticipation variability. We conclude that HD ity is smaller than that between cells (Blair et al. 1997; Taube
afferent dynamics and the statistics of rat head movements are important and Muller 1998). Additionally, estimates of the mean ATI
in generating HD anticipation. This result contributes to understanding (recorded from the same brain area, but across cells and
the functional circuitry of the HD system and has methodological impli- subjects) can differ by 25–30 ms between different studies
cations for studies of HD anticipation. (comparing Blair et al. 1998 and Stackman and Taube 1998)
including studies from the same group (comparing Bassett et
INTRODUCTION al. 2005 and Taube and Muller 1998). Finally, a recent report
(Bassett et al. 2005) showed an ATI increase of 40 ms when
To support complex behaviors, the brain must convert sen- rats were passively rotated compared with freely moving.
sory information into abstract and persistent representations. Where examined, such ATI differences could not be explained
The rodent head-direction (HD) system (Ranck 1984; Sharp et by behavioral variables such as average turning velocity or
al. 2001a; Taube et al. 1990a,b; Wiener and Taube 2005) is a turning bias (Bassett et al. 2005; Blair et al. 1997). Thus
striking example of a “cognitive” representation without a although individual cells to some extent have a characteristic
direct sensory correlate. A given HD cell is maximally active ATI, there are also unexplained differences.
when the animal’s head is facing that cell’s preferred firing Previous theoretical models exploring the possible origin of HD
direction in the horizontal plane, irrespective of location or anticipation (Goodridge and Touretzky 2000; Redish et al. 1996;
ongoing behaviors (Taube et al. 1990a). Different cells have Xie et al. 2002) do not account for these variations. Additionally,
different preferred directions, evenly covering the directional they fall short of generating the long ATIs observed in the lateral
space to form a “neural compass” representing the animal’s mammillary nuclei (LMN): simulations reported in Goodridge
current HD (Baird et al. 2001; Johnson et al. 2005). The and Touretzky (2000) only resulted in 30 ms of anticipation in the
compass can be updated by visual inputs, yet it persists in most favorable case, whereas mean experimental values range up
darkness (Blair and Sharp 1996; Goodridge and Taube 1995; to 67 ms (Stackman and Taube 1998).
but see Chen et al. 1994; Mizumori and Williams 1993). The In this modeling study, we present a novel hypothesis for the
generation of anticipation in the HD system. Like authors of
Present address and address for reprint requests and other correspondence:
M.A.A. van der Meer, Dept. of Neuroscience, University of Minnesota, 6-145 The costs of publication of this article were defrayed in part by the payment
Jackson Hall, 321 Church St. SE, Minneapolis, MN 55455 (E-mail: of page charges. The article must therefore be hereby marked “advertisement”
vande642@[Link]). in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.

[Link] 0022-3077/07 $8.00 Copyright © 2007 The American Physiological Society 1883
1884 VAN DER MEER, KNIERIM, YOGANARASIMHA, WOOD, AND VAN ROSSUM

A
1 2 left 4 5
MVN
tracking unit
update
data from gain
HD
moving rat 3 γ ring

input right modified output

angular head velocity MVN angular head velocity
unit

B
adaptation rebound
in
τ
A τ
R
out

C left right
E left right left
120 120 30
AHV (º/s)

actual HD
input

head direction (deg)

Downloaded from [Link] on May 9, 2008

model output
shifted output
0 2 0 2
D adaptation
MVN activity

rebound
0 2
MVN activity

0 2 0 2
-30 time (s)
time (s) time (s)
FIG. 1. Adaptation and rebound generate anticipation in the model, illustrated here with an artificial angular head velocity (AHV) input pattern. A: schematic
model layout. Tracking data from freely moving rats (1) yields an AHV input pattern (2), which is split up into its left (clockwise) and right (counterclockwise)
components. The 2 AHV components are the input to corresponding cross-inhibiting medial vestibular nucleus units (3), which filter the signal by means of
adaptation and postinhibitory rebound (B). The filtered signal from the medial vestibular nucleus (MVN) units (4, dark gray line; black line is the original input
signal) is used to update a ring attractor network representing head direction (5). The gain ␥ controls how much the head-direction (HD) network is updated in
response to input, and is chosen to minimize tracking error (METHODS). B: schematic representation of adaptation and rebound in response to a step input.
Adaptation and rebound are parameterized by their respective strengths (A, R) and time constants ␶. C: example artificial “left-right-left” input AHV profile. D:
adaptation (open squares) and rebound (gray circles) currents are added to the AHV input pattern to yield the net activity (black triangles) of the MVN neurons,
which update the HD module. The 2nd “left” pulse results in a bigger response than the identical 1st one, due to rebound generated by the right input. E: model
output. Note how during the initial stage of a turn, the model output (black circles) is updated faster than the actual HD (open squares). To find the resulting
anticipatory time interval (ATI), the model output is shifted in time until the error with respect to actual HD is minimal (but nonzero). In this example, the output
was shifted by 48 ms (gray triangles).

previous models (e.g., Redish et al. 1996; Skaggs et al. 1995; ment pattern, thus providing an explanation for the observed
Song and Wang 2005), we assume that the HD system inte- ATI variability as well as way to test the model directly against
grates angular head velocity (AHV) from the vestibular sys- experimental ATI values. In support of our hypothesis, we first
tem. However, we consider the firing dynamics of vestibular show that rats exhibit variations in the statistics of their head
neurons such that instead of being updated by a perfect ves- movements. ATIs generated by the model on these head move-
tibular AHV signal, the HD system effectively receives a ments correlate strongly with experimental values, accounting for
high-pass filtered version of that signal, which results in over half of the experimentally observed ATI variability.
anticipation. This filtering is achieved by incorporating the
effects of vestibular spike rate adaptation (a decreasing firing METHODS
rate response to a constant persistent stimulus) and postinhibi- Ring attractor network model
tory rebound firing (a transient increase in firing rate following Our model of the HD system uses a “ring” attractor network of
release from inhibitory input) as reported in Sekirnjak and du nonlinear units with generic rate-based dynamics, similar to previous
Lac (2002). We show that for physiological amounts of ves- models (Redish et al. 1996; Skaggs et al. 1995; Trappenberg 2002;
tibular adaptation and rebound, the model generates realistic Zhang 1996). However, the critical component of the model lies in the
ATIs on a large set of rat tracking data. Critically, the resulting dynamics of the input signal to this network (described in Input
anticipation depends on the statistics of the input head move- dynamics). Figure 1A shows a schematic of the complete model, of

J Neurophysiol • VOL 98 • OCTOBER 2007 • [Link]

 HD ANTICIPATION MODEL 1885

which the ring attractor network is the last processing stage. In TABLE 1. Model parameters used for the results presented, unless
general, like previous models, the ring network integrates AHV input stated otherwise
to yield a persistent representation of HD. Briefly, the network units
(representing populations of neurons) are placed on a ring, where Symbol Parameter Value
angular position on the ring corresponds to the unit’s preferred firing
direction. A rotationally symmetric matrix of recurrent weights is ␶ Single unit time constant 10 ms
constructed so that a subset of units is persistently active, forming a ␶A Input adaptation current time constant 200 ms
␶R Input rebound current time constant 200 ms
Gaussian-shaped activity packet, or attractor state, even in the absence dt Simulation time step 1 ms
of input (Amari 1977). The packet is stable at any position along the N Number of neurons in ring 100
ring, and its position corresponds to the HD encoded by the network. ␤ Activation function slope 0.07
The packet is moved around the ring by external inputs through wE Recurrent weights scale factor 15
addition of an asymmetric component to the weight matrix propor- ␴ Recurrent weight profile width 21.6 deg
wI Recurrent weights global inhibition level 9
tional to the magnitude of the input (Zhang 1996).
A MVN unit adaption 0.4
Specifically, following Stringer et al. (2002), the activity level u(t) R MVN unit rebound 0.4
of unit i in the HD ring is modeled by

␶
dui 共t兲
dt
⫽ ⫺ui 共t)⫹ 冘 j
wij (t)F[uj (t)]
Visual input
The attractor network described above provides a way for the HD
representation to be updated by idiothetic (self-motion) information
where ␶ is the time-constant of the unit and the activation function F from the vestibular system. However, the HD system can also be
is a sigmoid updated by visual information (Blair and Sharp 1996; Goodridge and

Downloaded from [Link] on May 9, 2008

Taube 1995; Taube et al. 1990b; Zugaro et al. 2003). Such visual
F共u兲 ⫽ 1/共1 ⫹ e⫺␤ u )
“fixes,” where the animal uses visual cues in the environment to set its
HD representation, could affect anticipation. To quantify this, we
with a slope denoted by ␤. The weight matrix w of the recurrent model visual fixes as an instantaneous resetting of the HD represen-
connections in the network consists of a constant symmetric compo- tation in the network to the animal’s actual HD. Specifically, at the
nent ws, and two variable asymmetric components w{l,r}
a
, which move time of a fix, the activity packet is uniformly moved such that the
the activity packet left (clockwise) and right (counterclockwise), absolute angular difference between the model and actual HD is
minimized.
respectively.

w共t兲 ⫽ w s ⫹ ␥ 关vl 共t)wal ⫹vr 共t)war 兴 (1) Input dynamics

The core of the model describes how the input units that update the
The symmetric weight matrix ws implements a standard local excita-
HD representation respond to the AHV ␻(t) to be tracked by the
tion and global inhibition connectivity. The strength of the recurrent model. The AHV signal is split up into “left” and “right” components,
excitation depends on the difference angle between the preferred which feed into cross-inhibiting left (vl ) and right (vr ) units, consis-
directions of the units tent with vestibular system physiology (Markham et al. 1978;
Shimazu and Precht 1966). These units correspond to a population of
⫺d ij2
AHV-sensitive neurons afferent to the HD system (see DISCUSSION).
w sij ⫽ wEe 2 ␴ 2 ⫺ wI (2)
The firing properties of the input units are augmented with adaptation
where dij is the angular difference between the preferred firing and rebound firing, as described in MVN neurons by Sekirnjak and du
Lac (2002); for clarity, we refer to the input units as “MVN units”.
directions of the units i, j; ␴ denotes the width of the Gaussian weight
This arrangement is shown schematically in Fig. 1A. MVN unit
profile, and wE and wI determine the strength of the excitation and
activity equals the AHV minus an adaptation current (IA), plus a
inhibition, respectively. These are chosen such that the network has a rebound current (IR), as follows
stable attractor state.
The asymmetric components in Eq. 1 are obtained by multiplying v r(t) ⫽ 关␻r 共t) ⫺ I rA共t兲 ⫹ I rR共t兲 ⫺ vl 共t)]⫹ (3)
a
the asymmetric weight profiles w{l,r} with the activity of the vestib-
ular inputs v{l,r} and a constant gain (scaling) factor ␥ (see section v l共t兲 ⫽ 关 ␻ l共t兲 ⫺ I lA共t兲 ⫹ I lR共t兲 ⫺ v r共t兲兴 ⫹
a
Input dynamics). For the asymmetric weights w{l,r} , we use the The last term describes vestibular cross-inhibition from the contralat-
derivative of the symmetric weights given by Eq. 2, as in previous eral MVN unit. The angular velocity inputs ␻l(t) [and ␻r(t)] are
models (Stringer et al. 2002; Zhang 1996) simply the right (left) components of the input angular velocity signal,
i.e., ␻l(t) ⫽ [␻(t)]⫹ and ␻r(t) ⫽ [⫺␻(t)]⫹, where [x]⫹ ⫽ max(x,0).
dij ⫺dij2 Prolonged activation of a MVN unit activates an adaptive current
共w al 兲 ij ⫽ e2␴2 IA, illustrated in Fig. 1B. The adaptive current is linear in the AHV
␴2
input and is modeled with a first-order differential equation. For the
for the left-turning weights and war ⫽ ⫺wal for the right-turning left unit
weights.
dI lA
In summary, rotation of HD is effectively implemented through a ␶A ⫽ ⫺IAl ⫹ A␻l 共t) (4)
modulation of the weights. We chose this simple implementation dt
because the details of the update mechanism are not relevant for this and similar for the right unit. The parameter ␶A describes the time
study, and the actual biological mechanism is not known, although constant of the adaptation build-up, and the parameter A denotes the
more detailed, biologically plausible, update mechanisms have been adaptation strength. The adaptation current was chosen to depend on
suggested (Boucheny et al. 2005; Compte et al. 2000; Goodridge and the input rather than the activity of the unit itself to allow straight-
Touretzky 2000; Song and Wang 2005). forward interpretation of the parameter A as a fraction of the input, in

J Neurophysiol • VOL 98 • OCTOBER 2007 • [Link]

1886 VAN DER MEER, KNIERIM, YOGANARASIMHA, WOOD, AND VAN ROSSUM

line with Sekirnjak and du Lac (2002). This does not change the model on the set of tracking data with a fixed gain equal to the mean
qualitative features of the model. A parameter value A ⫽ 0 corre- gain previously obtained for each input pattern separately using the
sponds to no adaptation and A ⫽ 0.4 to a steady-state firing rate of error-minimizing procedure described in the preceding text. Mean
60% of initial activity. ATI values obtained this way did not differ significantly from the
The postinhibitory rebound currents I lR (and I rR, illustrated in Fig. values reported, although both tracking error and ATI variability were
1B) build up due to cross-inhibition from the contralateral side. The higher.
rebound current has a time constant ␶R and a net rebound gain factor Using the model’s population vector output to compute ATIs is
R, which combines the strength of the contralateral inhibition and much more precise than the single-cell-based method used in the
amount of rebound of the MVN unit experimental literature because all units in the model are accessible.
To obtain the ATI of single cells in the model, the time shift which
I lR maximized the mutual information between the cell’s activity and the
␶R ⫽ ⫺I lR ⫹ R␻r(t) (5)
dt HD input was calculated, as was done for HD recording data (Taube
and Muller 1998).
and vice versa for I rR. Thus R ⫽ 0.2 means that when the left input ceases
to be active after a long time of activity (t ⬎⬎ ␶R), the right MVN unit
rebounds with an initial activity level of 20% of the left input (cf. Eq. 3). Poisson HD cell simulation
For shorter rotation times, the rebound current is less. In Fig. 1, the
To assess the anticipation variability that can be expected from
adaptation and rebound currents are shown for a simple velocity profile.
spiking variability alone (Fig. 3C), we used the fact that HD cell firing
is thought to be approximately Poisson (Blair et al. 1998). We
Choice of parameters simulated a Poisson HD cell where the probability of a spike at each
simulation time step (1 ms) was defined by a Gaussian tuning curve
For all simulations, we used A ⫽ R ⫽ 0.4 and ␶A ⫽ ␶R ⫽ 200 ms (␴ ⫽ 21°) for direction, with the mean ⫾5° different for clockwise
unless stated otherwise. A ⫽ 0.4 corresponds to a strongly adapting

Downloaded from [Link] on May 9, 2008

versus counterclockwise turns respectively, to generate anticipation.
MVN neuron in Sekirnjak and du Lac (2002). The time constants were The cell has a 100-Hz peak firing rate, 20-Hz mean firing rate, and 40
extracted from a single-exponential fit to the published data. The ms of anticipation on average. The simulation used a 0.1-Hz sinusoi-
physiological value of R is more difficult to estimate using currently dal directional profile 8 min in length, covering the entire directional
available data because it depends on both the characteristics of the range. For comparison with the experimental data from Blair et al.
MVN neuron and on the strength of the cross-inhibition it receives. (1997), we used their method for computing ATIs in this simulation:
To assess the effect of the simplifications and parameter choices of two tuning curves for clockwise and counterclockwise turns were
our input implementation, we also simulated the MVN units v using constructed, and the HD cell spikes were shifted in time until the two
the two conductance-based spiking MVN neuron models described in curves overlap (Blair and Sharp 1995).
Sekirnjak and du Lac (2002). Although we were unable to exactly
reproduce background firing rate of the reported model, the rate
dynamics matched closely those reported there. To accommodate Tracking data
these neuron models into our HD network, the left model neuron To provide the model with realistic inputs, rat HD tracking data
received excitatory input current proportional to ␻l and inhibitory were used. Tracking data were obtained from adult male Long-
input proportional to ␻r (and vice versa for the otherwise identical Evans rats (n ⫽ 3) running a plus maze ⬃1.8 m in diameter during
right model neuron). The neurons’ spiking output was convolved with the Neural Systems and Behavior summer course at the Marine
a Gaussian of 100 ms SD to serve as input v to our rate model. As the Biological Laboratory, Woods Hole, MA. A color camera mounted
precise proportionality between AHV and input current is not known, overhead sampled the light-emitting diode (LED) pattern of a
we have chosen values which make use of the full dynamic range of HS-54 recording headstage (Neuralynx, Tucson, AZ) and a 10-cm-
the neurons: we used 100 nA s/° (i.e., the instantaneous net input long boom attached to the rats’ head at 30Hz, as described in
current to the neuron is 100 nA per unit angular velocity, which is in Yoganarasimha et al. (2006). The effect of missing samples was
degrees per second) for excitation and 1,000 nA s/° for inhibition. minimized first by only accepting recording sessions where ⱖ90%
This corresponds to strong inhibition and thus strong rebound; hence, of the tracking samples were sufficient to determine instantaneous
the rebound simulation results should be viewed as an upper bound on HD at the sampling points and second by extracting AHV from
how much anticipation this mechanism generates. pairs of consecutive nonmissing samples only and using those to
reconstruct a continuous HD profile. In other words, only AHV
Computing ATIs information known to sampling precision was used. This procedure
left 2 h and 22 min of usable data.
The HD encoded by the model is extracted using a linear population Because rat HD movements are sampled with finite precision and at
vector read-out (the vector sum of the preferred firing directions of the too coarse a time scale to be used as model inputs directly, we
model cells weighted by their firing rate) (Georgopoulos et al. 1982). approximated the rat’s true HD profile by fitting three different cubic
To accurately track AHV input, the gain ␥ needs to be set. The gain splines to each dataset: a spline that is forced to pass through all
determines by how much the HD representation is updated in response sampled points (natural or variational spline; no smoothing), as well
to a given input. An incorrect gain value leads to build-up of error in as a “weak” and a “strong” smoothing spline (MATLAB Spline
the HD representation over time and can lead to artifactual anticipa- Toolbox function spaps, tolerances 0.001 and 0.002, respectively),
tion or lag. For each HD input profile [e.g., a 60-s segment of tracking which need not pass through every sample but are constrained by an
data (see following text), or an artificial step pattern], the optimal gain error function instead. The type of smoothing/interpolation applied to
is found by minimizing the error between the model’s population the data slightly affects the absolute ATIs reported here but not the
vector output and the time-shifted actual HD. The error is defined as qualitative pattern of the results. All simulations are done on the
the summed absolute difference per second between input pattern and weakly smoothed data.
decoded model output; using a sum of squares error yielded compa- We randomly selected 100 2-s- and 30 nonoverlapping 60-s-long
rable results on a representative subset of the data. Both gain and time segments that had a mean absolute angular velocity between 45 and
shift were varied to minimize the error; the time shift corresponding 1,000°/s. The same set of 60-s segments was used for all tracking data
to the minimal error output is the ATI for that pattern. To assess the simulations except for the spiking models from Sekirnjak and du Lac
effect of computing the gain for each input separately, we also ran the (2002), for which the 2-s set was used to reduce simulation time. The

J Neurophysiol • VOL 98 • OCTOBER 2007 • [Link]

 HD ANTICIPATION MODEL 1887

A B
200 -A-R 200 -A-R
+A-R +A-R

AHV

AHV
160 -A+R -A+R
160
+A+R +A+R

120 120
ATI (ms)

80 80

40 40

0 0
0 1 2 3
10 10 10 10 0.1 1 5
AHV (deg/s) input frequency (Hz)
FIG. 2. Anticipation in the model depends on input frequency but not on input magnitude. A: ATI as a function of AHV using a step input. Adaptation (A)

Downloaded from [Link] on May 9, 2008

and rebound (R) can both generate anticipation independently (⫹A⫺R/gray diamonds and ⫺A⫹R/gray triangles, respectively), and add supralinearly
(⫹A⫹R/black squares). B: ATI as a function of frequency of a sinusoidal input. Higher input frequencies result in less anticipation. At low frequencies, adaptation
generates more anticipation than rebound. For analytical results supporting these simulations, see the APPENDIX. The step input patterns were 6 s long: 1 s of
constant input followed by 5 s of rest. The sinusoidal input patterns were 10 s long, the ATI was computed over the last 5 s to eliminate the effects of initial
transients. Parameters in METHODS.

mean ⫾ SD absolute AHV was 86.4 ⫾ 8.5°/s on the 60-s data sessions were obtained by averaging across the segments of that
(smallest: 67.0°/s, largest: 97.9°/s) and 100.5 ⫾ 29.3°/s on the 2-s data session.
(smallest: 49.6°/s, largest: 179.6°/s). All input tracking data were
arbitrarily assumed to start at zero HD; that is, the activity packet was
RESULTS
always initialized at the same point in the ring. Because of the
rotational symmetry of the system, this does not affect the generality We implement a single-layer continuous “ring” attractor
of the results. network model of the rodent HD system, similar to previous
models (Boucheny et al. 2005; Goodridge and Touretzky 2000;
Experimental anticipation data Redish et al. 1996; Song and Wang 2005; Stringer et al. 2002;
Xie et al. 2002; Zhang 1996). This type of model has a
To compare the model’s predictions directly to experimental HD
cell data, we used two recording data sets from two different behav-
continuum of stable states in which a subset of HD cells
ioral tasks. The first set is a subset of the data described in Yoga- representing the animal’s current directional heading is persis-
narasimha et al. (2006). Rats ran on an elevated circular track while tently active. To track the animal’s movement, the position of
HD cell activity from the anterodorsal thalamic nucleus (ADN) was the packet needs to be updated. In contrast to earlier models
recorded and their head movements tracked. We used the “baseline” and in accordance with physiology, we incorporate spike rate
sessions, where apparatus and cues are always in the same, stable adaptation and postinhibitory rebound of MVN neurons
configuration. The second set is of rats foraging for randomly scat- (Sekirnjak and du Lac 2002) (Fig. 1B) into the update mech-
tered chocolate sprinkles in a walled square box with a prominent anism. MVN neurons are thought to provide the vestibular
polarizing cue card, using the same recording procedure. This exper- input to the HD system (Stackman and Taube 1997). As a
iment is similar to recording conditions in previously reported HD result, the input signal integrated by the model HD network is
anticipation data. A smoothing spline was fitted to the tracking data as
in the preceding text.
a modified version of the actual AHV. Following vestibular
The first “circular track” dataset included 12 3- to 6-min-long anatomy (Markham et al. 1978; Shimazu and Precht 1966), the
recording sessions from a total of two animals, the second “square model contains left and right input units, the activity of which
box” set consisted of 9 9- to 12-min-long sessions from three animals provides the update signal for the HD “ring.” These units,
(1 of which also contributed to the circle dataset). In one animal in the representing populations of MVN neurons, increase their ac-
square box data set, the recording electrode was identified as being in tivity during a left and a right turn, respectively. They adapt in
the anteroventral thalamus (in the border region between the AV and response to constant input and cross-inhibit each other, result-
VA nuclei). Data from this animal did not appear different in firing or ing in postinhibitory rebound when a turn stops and inhibition
anticipation properties and were therefore included. Every recording is released. Adaptation and rebound are characterized by their
session was split up into 60-s segments, such that segment 1 for that strengths (A, R) and time constants (␶A, ␶R); we fitted these to
session was the first 60 s, segment 2 the second 60 s, and so on. The
ATI for each segment was then computed using the mutual informa-
the data in Sekirnjak and du Lac (2002) (METHODS). A sche-
tion method described in Taube and Muller (1998). Segments ⬍45 s matic diagram of the model layout is shown in Fig. 1A.
long, or containing ⬍400 spikes from a HD cell, were rejected. For The effect of MVN adaptation and rebound on our model
comparison with model output, when multiple cells were recorded HD system is illustrated in Fig. 1, C–E, where a model rat
simultaneously during a segment, the experimental ATI was taken to moves its head in an artificial left-right-left pattern. The AHV
be the mean of individual cells’ ATIs. ATIs for complete recording input to be tracked by the model is first split up into left
J Neurophysiol • VOL 98 • OCTOBER 2007 • [Link]
1888 VAN DER MEER, KNIERIM, YOGANARASIMHA, WOOD, AND VAN ROSSUM

A 120 B
rebound = 0
rebound = 0.4 120
100 neuron A
rebound = 0.8
100 neuron B
80
ATI (ms) 80

ATI (ms)
60
60
40
40
20
20
0
0
0 0.1 0.2 0.3 0.4 +A-R +A+R -A-R
adaptation
C D error (deg/s) per ms anticipation
20
0.4 0.3+

15
ATI std. dev. (ms)

0.3 0.25

Downloaded from [Link] on May 9, 2008

rebound
10
0.2 0.2

5 0.1 0.15

0 0 0.1
simul. data data model model 0 0.1 0.2 0.3 0.4
Poisson within all -A-R +A+R adaptation
FIG. 3. The model produces realistic levels of anticipation using rat tracking data input. A: mean ATIs (⫾SE) for different values of adaptation and rebound,
simulated over a set of 30 60-s-long tracking segments. Experimental estimates for the mean ATI are 39 – 67 ms. B: mean ATIs (⫾SE) using the
conductance-based vestibular neuron models (their neurons A and B) from Sekirnjak and du Lac (2002), simulated using 100 2-s tracking segments. The resulting
ATIs for adaptation only (⫹A⫺R), and both adaptation and rebound (⫹A⫹R) are a good match with the results in A. C: adaptation and rebound can account
for the experimentally observed ATI variability. Shown is the ATI SD for a simulated HD cell with Poisson spiking, along with 2 estimates of the experimental
value [black bars, experimental values calculated from the data in Blair et al. (1997), see main text]. Poisson spiking accounts for less variability than that shown
experimentally both within cells (data within) and combined within ⫹ between cells (data all). In the model, adaptation and rebound result in a level of variability
consistent with the experimental values when adaptation and rebound are included (white bars). D: matching adaptation and rebound parameters reduces tracking
error. Shown is a contour plot of tracking error divided by ATI on the set of 60-s tracking segments, where white corresponds to small error and black to large
error.

(clockwise) and right (counterclockwise) components (Fig. sentation can be seen in Fig. 1E. As the movement stimulus
1C). This input is then modified by adaptation (Fig. 1D, open starts, the model output (black circles) initially updates faster
squares) and rebound currents (Fig. 1D, gray circles) to yield than the actual HD (open squares). When the stimulus changes
the net MVN activity (Fig. 1D, black triangles). Specifically, direction, rebound provides an additional update boost. These
the left MVN unit responds to the constant leftward input and mechanisms effectively implement a high-pass filter or angular
adapts, whereas rebound current builds up in the right MVN acceleration component (Song and Wang 2005; Zhang 1996),
unit. The rebound becomes active once left movement stops although it cannot simply be described as a linear filter (see the
(releasing cross-inhibition), coinciding with the start of the APPENDIX). The result is anticipation: the model output precedes
right turn. The right MVN response to the right turn is boosted the actual HD in time. To calculate by how much the model
by the rebound current. Meanwhile, as the right unit adapts, anticipates, the model output is shifted in time to minimize the
rebound builds up in the left MVN unit. This rebound then error with respect to the actual HD (METHODS). The time shift
boosts the response to the following left turn: compare the corresponding to the minimal error (gray triangles in Fig. 1E)
response to the two identical turns in the left MVN unit. The is the model’s ATI. For this example, the model’s ATI was 48
second response is larger due to the rebound current generated ms.
by the intervening right turn.
The net activity of the left and right MVN units described in Anticipation on artificial input patterns
the preceding text moves the HD representation in the corre-
sponding direction, allowing the system to track a given input To characterize the model’s behavior further, we first
pattern. Exactly how much the HD activity packet moves in present simple artificial AHV patterns to the model. On step
response to MVN activity is determined by a fixed gain inputs of constant AHV (Fig. 2A), adaptation only (⫹A⫺R)
parameter ␥, which is chosen to minimize tracking error and rebound only (⫺A⫹R) both generate anticipation. When
(METHODS). For our example, the model’s resulting HD repre- adaptation and rebound are combined (⫹A⫹R), the resulting
J Neurophysiol • VOL 98 • OCTOBER 2007 • [Link]
 HD ANTICIPATION MODEL 1889

ATI is larger than the sum of the two separately. The ATIs when adaptation and rebound are matched and ⬃0.2 (Fig. 3D).
shown here can be derived analytically (APPENDIX): the theo- Sekirnjak and du Lac (2002) found that adaptation and rebound
retical values (⫹A⫹R: 133 ms, ⫹A⫺R: 66 ms, ⫺A⫹R: 40 ms, of a given cell were strongly correlated, suggesting that MVN
⫺A⫺R: 0 ms) match the simulations well. cells may be tuned to produce anticipation at the lowest
For sinusoidal input patterns (Fig. 2B), there is a strong possible tracking error. Additionally, we found that the track-
effect of input signal frequency on anticipation. Some intuition ing error depends on the time constant of adaptation and
about this result comes from the observation that anticipation rebound. The error per unit anticipation was minimal for a time
results from a transient boost at the start of the input signal. constant of ⬃175 ms when tested on the tracking data (not
Because this boost doesn’t happen instantaneously, it will only shown), matching the physiological time constants of adapta-
generate anticipation if the direction of the boost is consistent tion and rebound (Sekirnjak and du Lac 2002). This provides
with the subsequent signal; if the signal changes too fast, the further evidence that the system may be tuned to reduce
boost will be in the wrong direction and will not constitute
tracking error while generating anticipation.
anticipation. Comparing the model’s ATI on sinusoidal inputs
at low frequencies to that on step inputs, adaptation makes an
even bigger contribution to anticipation than rebound. This Between-session anticipation variability
occurs because at low frequencies, rebound decays before it
becomes active; at higher frequencies, like adaptation, it be- Using realistic data, the model produces ATIs comparable
comes ineffective because the signal changes too quickly. The with those observed experimentally. However, a salient aspect
theory provided in the APPENDIX provides a good description of of the experimental data is its variability: when a given HD cell
the simulation results but for periodic stimuli is limited to the is recorded during multiple sessions, the resulting ATIs can
case where adaptation and rebound strength are equal (A ⫽ R).

Downloaded from [Link] on May 9, 2008

differ by up to 50 ms (although this within-cell variability is
reported to be on average less than that between different HD
Anticipation on realistic input patterns cells) (Blair et al. 1997; Taube and Muller 1998). We first
asked how much of the experimentally observed between-
Simulations on artificial input patterns show that in the session variability could be expected from random variations in
model, anticipation depends on the frequency of the input. Do the spike times of a HD cell. For direct comparison with the
behaving rats move their heads at frequencies capable of
data in Blair et al. (1997), we simulated a HD cell with Poisson
supporting the experimentally observed amounts of anticipa-
firing over repeated runs of the same AHV profile, 8 min in
tion, for realistic parameters of the model? To address this
question, we ran the model on 30 1-min-long input patterns, length. The resulting SD of the ATI was much lower than both
selected randomly from a large corpus of rat tracking data the between-session and overall SDs reported experimentally
(METHODS). The resulting ATIs are plotted in Fig. 3A for (Fig. 3C, black bars). Thus ATI variability cannot be explained
different values of adaptation and rebound. As expected, the by random variations in spike times.
ATI values on the tracking data are lower than the ATIs in To what extent does the model capture ATI variability?
response to step patterns (Fig. 2A) because of higher-frequency Without adaptation and rebound, the model yields little ATI
components present in real rat head movements. The mean variability, but when adaptation and rebound are included, the
adaptation value found experimentally appears to lie between ATI variability increases to experimentally observed levels
0.2 and 0.3 (Sekirnjak and du Lac 2002); the value for rebound (Fig. 3C, white bars). This occurs because the dynamics of
is more difficult to estimate because it does not depend solely adaptation and rebound interact with the statistics of the input
on the intrinsic firing properties of the neuron but also on the pattern as shown by the frequency-dependence of Fig. 2 and
strength of vestibular cross-inhibition. For the ATI in the most the APPENDIX. As a result, different input patterns result in
anticipatory HD area, the LMN, values of 39 and 67 ms have different amounts of anticipation. For realistic data, this idea is
been reported (Blair et al. 1998; Stackman and Taube 1998). illustrated in Fig. 4A where the input patterns generating the
Hence even if rebound were completely absent, the model can least (top) and most (bottom) anticipation are shown. As
still account for a substantial part of the observed anticipation. evidenced by their respective power spectra (Fig. 4B), the
An obvious concern is that the results reported here depend low-anticipation pattern has more power in higher frequencies
on simplifications in the model MVN neurons. To assess this, than the high anticipation pattern. The high-power, low-antic-
we used the spiking neuron models provided by Sekirnjak and ipation pattern has a larger tracking error (average tracking
du Lac (2002) as inputs to our HD network (METHODS), again error of 7.1°/s compared with 0.9°/s for the low-power pattern).
using actual tracking data as inputs. A literal implementation of Although in the limit, perfect tracking (no error) results in no
these spiking neurons coupled to our HD network yielded anticipation, not all errors necessarily result in more anticipa-
comparable ATIs (Fig. 3B), indicating that the anticipation is tion. For instance, the large overshoots in the model output (see
robust and does not depend on details or simplifications in our magnified inset of the high-power pattern), caused by the
model. adaptation and rebound currents, contribute little to ATIs but
In generating anticipation, the model necessarily introduces generate substantial tracking error. Apart from differences in
tracking error, as illustrated in Fig. 1. Perhaps the biophysical frequency content, the rat in the low-power pattern held its HD
properties of MVN neurons are optimized for generating an- constant for long periods; although this does not affect ATIs
ticipation while reducing tracking error. In the APPENDIX, we directly, in the dataset it tended to be the case that low-
show that for step profiles, tracking error is minimal when frequency head movements were accompanied by stationary
adaptation and rebound have equal strength. On the realistic periods. This illustrates that rats exhibit a range of movement
tracking data set, the error per unit anticipation is minimal patterns, which in the model result in different ATIs.
J Neurophysiol • VOL 98 • OCTOBER 2007 • [Link]
1890 VAN DER MEER, KNIERIM, YOGANARASIMHA, WOOD, AND VAN ROSSUM

model without [Fig. 5B, ANOVA F(1,29) ⫽ 12.75, P ⬍

0.001].

Influence of visual input

The HD system can update its representation using visual
inputs as evidenced by the system’s ability to recall a previous
orientation when returned to a familiar environment (Golob
and Taube 1997; Taube et al. 1990a). In darkness, HD cell
preferred firing directions exhibit drift (Goodridge and Taube
1995; Knierim et al. 1998), suggesting that visual updates help
maintain a consistent relationship between the animal’s actual
HD and the HD representation. To examine the effect of such
updates on anticipation, we assume that visual updates happen
at random (Poisson) time intervals and reset the HD represen-
tation to the correct value (METHODS). An example of this is
FIG. 4. Differences in the statistics of tracking data input result in different shown in Fig. 6A, where updates are indicated by the black
ATIs. A: input patterns resulting in the lowest (top) and highest (bottom) ATIs. dots, and the HD representation (black line) can be seen to
The input is shown in black with the model output overlaid in gray; the (top) “jump” to the actual HD (gray line). Figure 6 shows the
high-ATI pattern contains strong high-frequency components compared with
the (bottom) low-ATI pattern. B: angular head velocity power spectrum for the dependence of ATI on the frequency of visual updates: in

Downloaded from [Link] on May 9, 2008

5 lowest ATI patterns (top) and the 5 highest ATI patterns (bottom). Low ATI general, more frequent updates result in lower ATIs as would
patterns (top) have higher power at higher frequencies (note the peak at 1.5 Hz be expected from the limit where actual and model HD are
and the presence of ⱖ3 Hz components) than high ATI patterns (bottom). equal. Although how often rats actually update their HD
system in such a way is not known, at low update frequencies
Within-session anticipation variability (of ⬍1 Hz), ATIs are relatively unaffected.
We have shown that the model provides a possible expla-
nation for the observed between-session variability in HD cell Comparison with experimental data
anticipation by exploiting differences in the AHV input. How-
ever, it is also known that there is substantial within-session To test the idea that differences in AHV input contribute to HD
variability between simultaneously recorded HD cells (up to 30 anticipation directly, we used two recording data sets from dif-
ms) (Blair et al. 1997). To address this issue, we computed ferent behavioral tasks. In both sets, activity from HD cells in the
ATIs for the units in the model individually (METHODS). An anterior thalamus was recorded from freely moving rats as their
example of this is shown in Fig. 5A: the white dots indicate the head movements were tracked (METHODS). In the first data set, rats
ATIs of the individual units (indexed on the vertical axis; ATI ran clockwise on a elevated circular track (12 sessions total from
on the horizontal axis), superposed on their mutual information 2 rats, of 3– 6 min in length each) (data from Yoganarasimha et al.
content with respect to the HD input as a function of time shift. 2006). The second set consists of data from rats foraging for
Individual unit ATIs for this example range from 37 to 67 ms; randomly scattered food pellets in a square box enclosure (9
the differences are due to the fact that each cell has a limited sessions total from 3 rats, 9 –12 min in length each). We compared
tuning curve, and “sees” a different part of the input pattern. As the model’s predicted ATI to the experimental value on a session-
has been shown before, different input patterns lead to different by-session basis. For each session, we computed the experimental
amounts of anticipation. When averaged over the complete ATI (see METHODS) and the ATI predicted by the model from the
tracking data set, the model with adaptation and rebound rat’s head movements during that session. (Experimental ATI
results in significantly more within-session variability than the values are sensitive to the exact time stamping and synchroniza-

A 100 B
Imutual
12
0.8
*
ATI std. dev. (ms)
cell

0.7 10
0.6 8
0 0.5
-20 0 20 40 60 80 6
time shift (ms) 4
HD (deg)

180 2
0
-180 -A-R +A+R
0 time (s) 60
FIG. 5. Adaptation and rebound can explain different ATIs of simultaneously recorded HD cells. A: ATIs of individual units in the ring attractor (top) when
tracking an example head direction profile (bottom). Units are indexed from 1 to 100 according to their preferred firing directions (location in the ring) spanning
the directional range. ATIs are computed as the peak of the mutual information between that cell’s activity and head direction over a range of time shifts. The
shaded area shows the mutual information as a function of time shift with the location of the white circle indicating the maximum information, and thus the ATI,
for that unit. Only units with sufficient sampling (⬎5% of time spent in the cell’s directional range) are shown. In this example, ATIs of “simultaneously
recorded” individual units range from 37 to 67 ms. B: average SD of within-session ATIs over 30 patterns without (white bar) and with adaptation and rebound
(black bar). Adaptation and rebound cause increased within-session ATI variability.

J Neurophysiol • VOL 98 • OCTOBER 2007 • [Link]

 HD ANTICIPATION MODEL 1891

FIG. 6. Visual updates affect anticipation at high, but not low, update frequencies. A: example model behavior when updates (time indicated by black dots)
happen at an average frequency of once per second. The model output (black line) can be seen to “jump” to the actual HD (gray line) at the time of an update.
B: average ATI on the realistic dataset decreases with update frequency, although below frequencies of 1 update per second (note log scale) ATIs are relatively
unaffected.

tion of video and neural data. Although we have made every effort To illustrate how the model accounts for the data, we looked
to ensure that the values reported here are as close to the real value at within-session ATI differences. Recording sessions were

Downloaded from [Link] on May 9, 2008

as possible, small inaccuracies in the absolute values reported here split up in 1-min-long segments (with segment 1 the 1st minute
may still be present.) Figure 7A shows scatterplots of the model of a session, segment 2 the 2nd, and so on) and ATIs computed
and experimental ATIs for both data sets (top: circular track, for each segment separately. Starting with the circular track
bottom: square box), where each data point corresponds to one data set, we then plotted the average experimental and model
recording session. On both data sets, the model and experimental ATIs for each segment (Fig. 7B, top). Both the experimental
ATIs are strongly correlated (Pearson product-moment correla- and model data show a significant increase in anticipation with
tion, circular track: r ⫽ 0.79, P ⫽ 0.002, square box: r ⫽ 0.92, segment (linear regression, model: R ⫽ 0.98, F ⫽ 198.6, P ⬍
P ⬍ 0.001), indicating that the model explains ⬎60 and ⬎80% of 0.001, data: R ⫽ 0.87, F ⫽ 26.2, P ⬍ 0.01): the HD cells
the experimental ATI variance on the two data sets, respectively. anticipated more as the recording sessions progressed. The

A B C 200
150
Experimental ATI (ms)

Power (deg2 /(s2 × Hz))

r = 0.79, p = 0.002 Model: R = 0.98 (p < 0.001)

80
Data: R = 0.87 (p < 0.01)
60
100
ATI (ms)

segment 1
40
100
20
50
0
6
-20
0 0
0 20 40 60 80 1 2 3 4 5 6 0 2 4 6 8
150 300
Power (deg2 /(s2 × Hz))
Experimental ATI (ms)

r = 0.92, p < 0.001 Model: R = 0.20 (p = 0.19)

80
Data: R = 0.04 (p = 0.60)
60 100
ATI (ms)

200
40
20 50
100
0
0
-20
0
0 20 40 60 80 1 2 3 4 5 6 7 8 9 10 0 2 4 6 8
Model ATI (ms) Segment Frequency (Hz)
FIG. 7. The model explains a substantial amount of experimental anticipation variability. A: scatterplots of model against experimental ATIs across 2 data
sets (top: circular track, bottom: square box). Each data point corresponds to 1 recording session. For both sets, the model and experimental values are
significantly correlated. Statistics shown are Pearson’s product-moment correlation coefficient r and associated significance level p. B: model accounts for
systematic within-session changes in ATI on the circular track data set top. Recording sessions are split up into 60-s segments and the average ATI ⫾SE per
segment plotted. For both data and model (linear regression statistics shown), there is a significant increase in ATI as the session progresses. Bottom: on the square
box data set, there is no (linear) within-session effect on ATI (left) in either the model or experimental values. On this data set, the model overestimates the ATI
(see main text). C: on the circular track data, the increase in ATI over segments is accompanied by a corresponding change in the head movement power spectrum
(top). Shown are average angular head velocity power spectra for each segment, with the top (dark) line corresponding to the 1st segment and the bottom (light)
line to the last. Note that with increasing segment, there is less power overall and the peak shifts to lower frequencies. Bottom: on the square box data set, the
power spectra for the segments do not differ systematically. Parameters used: A ⫽ 0.3, R ⫽ 0.4.

J Neurophysiol • VOL 98 • OCTOBER 2007 • [Link]

1892 VAN DER MEER, KNIERIM, YOGANARASIMHA, WOOD, AND VAN ROSSUM

model explains this increase by changes in the frequency although some differences between the two data sets remain
content of rat’s head movements. Average power spectra for unexplained. We conclude that the firing properties of neurons
each segment are shown in Fig. 7C (top); the darkest line afferent to the HD system, specifically their dynamic response
corresponds to the first segment, with the line for each subse- to head movements, may be important in generating HD
quent segment colored progressively lighter. The first segment anticipation. This idea has implications for our understanding
(top black line) has higher overall power and a peak at a higher of the circuitry of the HD system as it stands in contrast to
frequency than the last segment (bottom, light gray line). As in previous proposals that treat anticipation as a consequence of
Figs. 2B and 4, higher frequency components result in less circuitry within the HD system. Furthermore, the result that
anticipation in the model and in the recording data. ATIs depend on the statistics of head movements has method-
In contrast, the square box data set did not show an effect of ological implications for the study of ATIs.
segment on ATI for either model or experimental data (Fig. 7B,
bottom; linear regression, model: R ⫽ 0.2, F ⫽ 2.0, P ⫽ 0.19; Origin of the HD update signal
data: R ⫽ 0.04, F ⫽ 1.3, P ⫽ 0.60). Consistent with this, the
average power spectra per segment (Fig. 7, bottom) do not The present model uses the firing properties of neurons in
show a clear progression as in the circular track data set. As the MVN to generate anticipation in the HD system. Several
Fig. 7, A and B (bottom) illustrates, although there is a strong lines of evidence indicate that the MVN are a likely origin of
correlation between the model and the data, the model over- vestibular inputs to the HD system (for a review, see Brown et
estimates the experimentally observed ATI on the square box al. 2002). Anatomically, the MVN project to the dorsal teg-
data set in absolute terms and underestimates its modulation. mental nuclei of Gudden (DTN), through the nucleus preposi-
The fact that the model does not always reproduce the exact tus hyperglossi (nPH) and possibly also directly (Brown et al.

Downloaded from [Link] on May 9, 2008

ATI values is not surprising. Apart from the potential influence 2005; Hayakawa and Zyo 1985; Liu et al. 1984). DTN, in turn,
of different visual update frequencies (Fig. 6), the model is reciprocally connected with the LMN (Gonzalo-Ruiz et al.
implements a simplified view of AHV input signal processing. 1992). Both these areas contain AHV as well as HD cells, and
In modeling MVN adaptation and rebound with a single cause loss of downstream directional firing when lesioned
exponential, it does not include contributions from mecha- (Bassett et al. 2007; Blair et al. 1998; Sharp et al. 2001b).
nisms with different time constants (e.g., adaptation in the Similarly, bilateral lesions of the vestibular apparatus itself
vestibular apparatus itself). Such contributions would respond abolish directional activity in HD areas (Stackman and Taube
to different components of the head movement power spec- 1997). Recording studies (albeit to our knowledge not in rats)
trum, potentially contributing differences between different have shown that semicircular canal-dependent MVN neuronal
data sets not seen in the current implementation. Also several activity contains AHV information (e.g., Fuchs and Kimm
experimental factors beyond the scope of the model may affect 1975; Melvill Jones and Milsum 1971; Vidal and Sans 2004),
absolute ATI values, including the use of different subjects for which is theoretically sufficient to update the HD representa-
the two data sets or the walls of the square box environment tion. Thus given that the vestibular AHV signal appears to be
restricting head movements artificially (e.g., the rat spends responsible for updating the HD system, its properties are
much time searching for food pellets along the walls or in the relevant to models of the HD system.
corners and thus does not have the freedom to move its head in We based our model on the intrinsic firing properties of
all directions). In our view, the fact that the model can account MVN neurons (Sekirnjak and du Lac 2002) as reported in rat
for much of the experimentally observed anticipation variance brain slices. In recordings from intact animals, the MVN
on both data sets without tuning of parameters is more com- response is reported to have a phase lead relative to a sinusoi-
pelling than if we had fit the parameters to the anticipation data dal AHV input (Fuchs and Kimm 1975; Melvill Jones and
without physiological justification or if we had invoked artifi- Milsum 1971; Vidal and Sans 2004), which decreases and even
cial mechanisms to produce a better fit. lags with increasing frequency (Kaufman et al. 2000). This
pattern is consistent with our model. Although it is possible
DISCUSSION
that this phase lead results from the properties of vestibular
afferents (e.g., Fernandez and Goldberg 1971) rather than, or in
We provide a novel hypothesis on how anticipation is addition to, intrinsic MVN firing properties, our implementa-
generated in the HD system. Following the in vitro firing tion captures the relevant dynamics.
properties of MVN neurons, which are thought to provide In principle, the model’s ability to generate anticipation only
AHV information to the HD system, we incorporate physio- depends on the HD input signal being a high-pass filtered AHV
logical levels of adaptation and rebound firing in our model. signal. More generally, any input signal that also contains a
The model produces realistic anticipation without further pa- time-derivative component (in the HD system, AHV plus an
rameter tuning when run on a corpus of rat tracking data. In the acceleration component) will result in anticipation when inte-
model, the statistics of the rat’s head movements interact with grated and decoded provided that the encoded quantity (HD)
the high-pass filtering generated by adaptation and rebound, varies smoothly (Puccini et al. 2007). The neural mechanisms
such that movement patterns with high-frequency components that accomplish this in the HD system need not exclusively be
result in lower anticipation than lower-frequency movements. implemented in the MVN. Because the precise update circuit is
We show that rats exhibit variations in the power spectrum of not yet known, we have concentrated on the possible contri-
their head movements in the relevant frequency range. When bution from the MVN for two reasons. First, it is a likely
the model output is compared directly against experimental source of input to the HD system as discussed in the preceding
recording data from two different behavioral tasks, the model text. Second, known physiological properties of MVN explain
accounts for 60 – 80% of the experimentally observed variance, a large fraction of the observed ATI. In this respect, it is
J Neurophysiol • VOL 98 • OCTOBER 2007 • [Link]
 HD ANTICIPATION MODEL 1893

notable that MVN firing properties appear to reduce tracking ATIs have been found. In the original proposal (Redish et al.
errors as much as possible. However, several other sources 1996), this population was placed in the postsubiculum (PoS),
could also contribute to HD anticipation. Apart from the the HD cells of which do not anticipate, and PoS sends a
characteristics of vestibular afferents mentioned in the preced- projection to LMN. However, a subsequent lesion study
ing text, the firing properties of AHV-sensitive neurons co- showed that PoS lesions left anticipation intact (Goodridge and
localized with HD cells could also be important. For instance, Taube 1997). This result does not doom the offset connection
different types of AHV neurons have been found in the DTN hypothesis as long another nonanticipating population (in DTN
(Bassett and Taube 2001b; Sharp et al. 2001) and have been perhaps) can provide the required connections to LMN. In
reported to exhibit complex activity before, during, and after contrast, the current model supports the view that those HD
head turns (Sharp et al. 2001). Fine-time-scale analysis of DTN cells closest to the source of the AHV signal should anticipate
and MVN AHV-sensitive neurons in rats, for instance during a most, with ATIs decreasing as the signal is propagated.
controlled head movement paradigm (Bassett et al. 2005;
Zugaro et al. 2001), would address whether these neurons Visual input and tracking error
could contribute to anticipatory firing. As suggested by previ-
ous models, specific connectivity patterns between HD cells Rats can use visual information to update their HD repre-
could also contribute to anticipation: this is discussed in the sentation (Taube et al. 1990b), and a brief view of familiar
following paragraph. landmarks improves homing performance in hamsters (Etienne
et al. 2000). If a visual update involves setting the represented
Relation to previous models HD to the animal’s true HD, the model’s ability to anticipate is
limited by the frequency of such updates. In the present model,

Downloaded from [Link] on May 9, 2008

A central issue concerning anticipation in the HD system is a noticeable reduction in anticipation is only seen at update
whether it is generated by circuitry within the HD system frequencies of ⬎1 Hz. The actual frequency at which visual
(perhaps as a consequence of the connections required for updates occur is not known, but there is a body of behavioral
angular path integration) or “inherited” from elsewhere. Zhang evidence that suggests that when animals are familiar with an
(1996) suggested a generic angular acceleration component in environment or task, they tend to adopt egocentric (praxic)
the input or update signal to account for anticipation in the HD over visual cue-based (piloting) strategies (Packard and Mc-
system, the origin and neural implementation of which is Gaugh 1996; Tolman et al. 1946). In such cases, the influence
unclear. The current model contains a physiologically plausible of visual inputs on anticipation would be expected to be small.
implementation of this idea. Other models have used offset However, in principle it is possible that differential update
connections between HD areas to generate anticipation (Goo- rates contribute to anticipation variability.
dridge and Touretzky 2000; Redish et al. 1996) for which there Our model of the visual update mechanism contains at least
is some indirect experimental evidence (Blair et al. 1997). two simplifications: it uses instantaneous updates, whereas
However, simulations of the offset scenario could only account experimental evidence indicates that although fast, they can be
for 30 ms of anticipation (Goodridge and Touretzky 2000) in expected to take at least ⬃100 ms (Zugaro et al. 2003). Also it
the most favorable case, whereas mean ATIs of 67 ms have seems likely that the update mechanism does not always update
been reported (Stackman and Taube 1998, note erratum). to the animal’s exact true HD: for instance, HD cell preferred
Offset connections and afferent anticipation/rebound are not firing directions shift ⬃5° when an animal is returned to a
mutually exclusive, and it is possible both contribute to antic- familiar environment (Taube 1995; Taube et al. 1990a). How-
ipation: the input dynamics of the current model could be ever, these mechanisms would not be expected to introduce a
combined with an additional offset layer. Such a combination bias under circumstances where head movements cover the
would explain both the within-cell variability of ATIs and the directional range uniformly. It is conceivable that with a
reported consistency of mean differences in ATI across indi- sufficient number of simultaneously recorded HD cells, the
vidual cells. Like offset connections, our model provides a way frequency and dynamics of visual updates could be examined
for the HD signal to anticipate yet still be sensory; this in more detail. A different possible experimental test for the
contrasts with earlier interpretations where anticipation was influence of visual inputs on HD anticipation could be to
interpreted as evidence against the HD signal being sensory compare ATIs under light and darkness conditions.
(e.g., Taube and Muller 1998). In contrast to earlier models, our model explicitly displays
Apart from providing additional anticipation, the current tracking error, resulting from the same adaptation and rebound
model differs from the offset connection explanation in two mechanisms that generate anticipation. This error could con-
important respects. First, the time constants of the mechanisms tribute to the drift of HD cell preferred firing directions
generating anticipation in the offset model are necessarily fast, observed in darkness (Goodridge and Taube 1995; Knierim et
on the order of the synaptic and membrane time constants al. 1998; Mizumori and Williams 1993). The model tracking
(5–20 ms). This implies that any frequency dependence of error is determined precisely by the AHV input, consistent with
anticipation will only become apparent at very high frequen- the systematic path integration errors found behaviorally
cies. In the current model with its slow (200 ms) dynamics, the (Séguinot et al. 1993), where particular return journey biases
effect of different input is clearly seen using naturally occur- result from specific, controlled sequences of turns.
ring variations in rat tracking data and can explain much of the
experimentally observed ATI variability. Second, because the Relation to experimental data
offset connection scheme generates anticipation in the receiv-
ing layer only, it predicts there should be a population of We report that there are variations in the power spectrum of
nonanticipating HD cells afferent to LMN, where the highest rats’ head movements which result in ATI differences in both data
J Neurophysiol • VOL 98 • OCTOBER 2007 • [Link]
1894 VAN DER MEER, KNIERIM, YOGANARASIMHA, WOOD, AND VAN ROSSUM

and model on two datasets from different behavioral tasks. Using explored with ATI measurements, is how the fact that ATIs in
a relatively simple model with only a single time constant, ADN are lower than those in LMN is to be reconciled with
60 – 80% of the experimentally observed anticipation variance can tuning curve deformations observed in ADN (Goodridge and
be explained, where none was explained before. This large cor- Touretzky 2000). Goodridge and Touretzky (2000) argue that
relation suggests a correspondence between the model and the if LMN drives ADN, as suggested by anatomy and lesion
actual ATI generation mechanism, although further experiments evidence, ADN would be expected to be more anticipatory
are required to ascertain to what extent this is achieved in MVN than LMN. The fact that this is not the case suggests a
versus possible contributions from other areas. contribution from the PoS input to ADN; in support of this,
However, when comparing the model’s performance on the Goodridge and Taube (1997) found ADN ATIs to be increased
circular track and the square box data set, it is clear that after PoS lesions. The critical question is whether the resulting
although there is a large correlation on both sets separately, and ATI is compatible with LMN driving ADN. For this, the ADN
there is a near-perfect match on the circular track data, the ATIs of PoS-lesioned animals would need to be compared with
model overestimates the ATI on the square box data. A LMN ATIs; as before, such a comparison could be confounded
possible factor contributing to this could be that on the circular by the influence of differences in head movements between the
track, HD cell tuning curves are observed to shift backward two groups.
relative to the animal’s running direction (Yu et al. 2006). We Some early interpretations of ATIs suggested that they could
believe this to be unlikely for two reasons. First, both the result from motor efference copy or proprioceptive feedback
model and experimental data ATI analysis are always run on (Taube 1998; Taube and Muller 1998). In contrast, the present
isolated 60-s-long segments of tracking data, where no as- model shows how simple transformations of a known sensory
sumptions about a cell’s preferred firing direction in that or input can generate anticipation without requiring specialized

Downloaded from [Link] on May 9, 2008

other segments are made. Second, even if there is a significant neural circuitry. Apart from the strong correlations between
preferred firing direction shift within a segment, this would model and data reported here, this idea is supported further by
only result in anticipation bias if this shift was correlated with the experimental result that ATIs were not abolished by passive
a change in turning bias (e.g., if in the 2nd half of a segment movement as predicted by a motor efference copy model
animals are more likely to turn right than in the 1st half). (Bassett et al. 2005). The present model exploits the fact that
Although there was an overall turning bias due to the nature of future HD can be extrapolated from current HD and AHV.
the task, no significant changes within segments were found, as Therefore it is unable to extract higher-order regularities or
would be expected from the 60-s marks not carrying any respond in anticipation of active movement. Although a direct
behavioral relevance. Additionally, there are clear, progressive functional role for HD anticipation remains to be demon-
changes in the animal’s head movements over time on this task strated, mechanisms like the one described here could gain
(Fig. 7C, top), which we independently show contribute to ATI valuable time for further processing.
differences (Figs. 2 and 4).
How then might the unexpectedly low ATI on the square box APPENDIX
task be explained? One possibility is a differential contribution
from visual updates, either using an “episodic” update mechanism To gain a deeper understanding of the model’s behavior and its
as explored in the present model or perhaps based on visual flow, precise dependence on the parameters, we study the anticipation
which can update the HD signal (Wiener et al. 2004). More analytically for two simple input profiles: a single step stimulus and a
sinusoidal input. In these cases, we obtain explicit expressions for
frequent visual updates in the relatively small, walled enclosure both the ATI and its associated tracking error.
compared with the open track (where visual cues are further
away) could result in a smaller ATI in the square box. Recordings
in darkness could help resolve this issue. Step input
First we consider an AHV profile that consists of a single turn over
Implications a certain angle. The input angular velocity has a step profile

Our results suggest it would be informative to examine to ␻ 共t) ⫽ ␻0 (0 ⬍ t ⬍ T )

what extent previously reported differences in anticipation can and ␻(t) ⫽ 0 otherwise. The actual angle of the head ␪0 is the integral
be explained by differences in head movements. Bassett et al. of ␻(t) and is given by ␪0(t) ⫽ ␻0t, for 0 ⬍ t ⬍ T, and ␪0(t ⬎ T) ⫽
(2005) report that during passive movement (rats rotated by an ␻0T, where without loss of generality we set ␪0(0 ⬎ T) ⫽ 0.
experimenter), HD cells anticipated significantly more than The input to the head-direction integrator is not the pure velocity
when the rats could move their heads freely. As suggested by signal, but one that is altered due to adaptation and rebound firing.
these authors, it seems likely that the active and passive Due to adaptation, the velocity input during the turn (given by Eq. 4
movement conditions had very different movement frequency in the main text) becomes
spectra; thus it is possible that the difference in anticipation can
v共t兲 ⫽ ␻0(1 ⫺ A ⫹ Ae⫺t/␶A) (0 ⬍ t ⬍ T)
be explained by a mechanism like the present model.
Similarly, in future studies of anticipation in the HD system, During the rotation the rebound current of the contra-unit builds up to
the potential contribution of differences in head movements a value ␻0R(1⫺e⫺T/␶R) . When the rotation stops, the velocity input
will need to be taken into account, for instance in comparing (from Eq. 5, main text) therefore behaves as
ATIs in light and darkness, it is possible that rats will make
v共t兲 ⫽ ⫺␻0R(1⫺e⫺T/␶R)e(T⫺t)/␶R (t ⬎ T )
very different head movements which could result in an ap-
parent ATI difference as a result of the light/darkness condi- Unlike our full simulation, we ignore subsequent dynamics and
tion. A different issue regarding HD circuitry, which could be processing of the integrator in what follows. Instead we assume that

J Neurophysiol • VOL 98 • OCTOBER 2007 • [Link]

 HD ANTICIPATION MODEL 1895

the velocity integrator integrates the input signal v perfectly with a A 1 ⫺t/␶
gain ␥. The estimated HD angle follows from integrating the input ␬ (t) ⫽ ␦(t) ⫺ e
1⫺A ␶
␪ 共t兲 ⫽ ␥␻0 关共1 ⫺ A)t ⫹ A␶A 共1 ⫺ e⫺r/␶A)] 共0 ⬍ t ⬍ T ) which captures both the transient overshoot when the velocity comes
⫺T/ ␶ A ⫺T/␶R (T⫺t)/␶R on and the undershoot from rebound when it turns off. To investigate
␪(t) ⫽ ␥␻0 兵关共I ⫺ A)T ⫹ A␶A 关I ⫺ e ] ⫺ R␶R (1⫺e )[1 ⫺ e ]} (t ⬎T)
how the model ATI depends on the properties of periodic input
The gain of the integrator is set such that when the integrator has patterns, we consider the simplified case of identical adaptation and
equilibrated, the estimated and actual angle are the same, rebound. The HD signal then is ␻共t兲 ⫽ ␻0e共2␲ift兲. We convolve this
limt3⬁␪(t) ⫽ ␻0T. This leads to input signal with the kernel

冕冋
t

册
␥⫽ ⫺T/␶A ⫺T/␶R
1 ⫺ A ⫹ A␶A /T(1 ⫺ e ) ⫺ R␶R/T(1 ⫺ e ) A
␻ * (t) ⫽ ␦共t ⫺ t⬘兲 ⫺ e⫺(t⫺t⬘)/␶ ␻0 e2 ␲ ift⬘dt⬘
␶共1 ⫺ A兲
In the limit T ⬎⬎ ␶A,␶R, and when A ⫽ R and ␶A ⫽ ␶R, the gain equals

冋 册
1/(1 ⫺ A). ⫺⬁
1 1 ⫺ 2A ⫹ 4␲2 共1 ⫺ A兲f 2 ␶2 ⫹ 2␲iAf␶
⫽ ␻共t)
1⫺A 1 ⫹ 4␲2 f 2␶2
Calculation of the ATI
From this the phase shift ⌬␾ follows as

冉 冊
The angular profile ␪(t) will show anticipation with respect to the
Im共␻*共0兲
true angle ␪0(t).To determine the amount of anticipation, we shift the ⌬ ␾ ⫽ atan
Re共␻*共0兲
true profile with an amount dt to minimize the error E(dt) ' 冕 ⬁
⫺⬁

冉 冊

Downloaded from [Link] on May 9, 2008

[␪0(t ⫹ dt) ⫺ ␪(t)]2dt. The ATI corresponds to the shift for which the 2 ␲ Af␶
⫽ atan
error is minimal, that is tATI ⫽ arg min E(dt).This minimization can 1 ⫺ 2A ⫹ 4␲2 共1 ⫺ A兲f 2␶2
be performed in the limit T ⬎⬎ ␶A,␶R. The resulting anticipation is For small anticipation (␾ ⬍⬍ 1) the time shift of the convolved
signal’s ATI is
1
t ATI ⫽ (A␶A ⫹ R␶R ) ⫹ O共1/T)
2共1 ⫺ A) ⌬␾ A␶
t ATI ⫽ ⫽ (A2)
where O(1/T) denotes correction terms of the order ␶A /Tand ␶R /T that 2␲f 1 ⫺ 2A ⫹ 4␲2 共1 ⫺ A兲f 2 ␶2
disappear as T increases. This equation fits our simulations in Fig. 2 The preceding expression shows that for low input frequencies f, the
very well. We see that both rebound and adaptation contribute to the ATI increases linearly with ␶ in accordance with the step input results.
ATI. Each contribution is proportional to its time constant. In the limit At higher frequency, however, the ATI decreases again. This decrease
when A, R ⬍⬍ 1, adaptation and rebound contribute equally to the ATI is faster in the case of large ␶. The maximum ATI occurs for 1/␶ ⫽ 2␲f
for these simple profiles.
冑1 ⫺ A/ 冑1 ⫺ 2A, with an ATI of tATI ⫽ {1/[2(1 ⫺ A)]}At. Thus a fast
time constant gives a small amount of anticipation which is less
Minimal error affected by fast-changing input, whereas a slow time constant gives a
lot of anticipation at slow frequencies, falling off rapidly (to below
Given that both adaptation and rebound contribute to the ATI, one fast ␶ levels) as input frequency increases.
can wonder how to best choose their contribution. This can be
evaluated by calculating the remaining error with the optimally shifted ACKNOWLEDGMENTS
profile, i.e., E(tATI). This yields
We thank M. Mauk, G. Rao, J. Siegel, and B. Kalmbach for assistance in
tracking data collection, M. Dutia and D. Perrett for discussion, and T. Trappen-
␻ 2
0 berg for freely available code that we used in the initial stages of this project.
E(t ATI) ⫽ [A␶A ⫺ R␶R ]2T ⫹ . . . (A1)
12共1 ⫺ A)2 Author contributions are as follows: concept (MvdM, MvR), code and data
analysis (MvdM), recording data collection (D.Y., J.K.), manuscript preparation
Importantly, the leading contribution to the error is proportional in the (MvdM with comments by MvR, J.K., and E.W.).
step duration T and hence diverges for long steps. To reduce the error
for a given ATI, adaptation and rebound should be matched such that GRANTS
A␶A ⫽ R␶R. When matched, the leading term disappears and the
sub-leading terms in T, indicated by the ellipsis in Eq. A1, become This work was supported by an Engineering and Physical Sciences Re-
search Council/Medical Research Council scholarship to M. van der Meer and
important. Assuming A ⫽ R and ␶A ⫽ ␶R, this term is E(tATI) ⫽ National Institute of Neurological Disorders and Stroke Grant R01 NS-39456
{1/[3(1 ⫺A)3]}␻20(A␶A)2[3␶A ⫹ 2A␶A]. To obtain a certain ATI, one to J. Knierim.
can still choose between a long time constant (␶A,␶R) or a large
strength (A, R). This expression shows that for the step profiles, it is REFERENCES
somewhat better to choose a short time constant as then the first term
in the brackets disappears, reducing the error. Amari S. Dynamics of pattern formation in lateral-inhibition type neural
In summary, adaptation and rebound contribute to the ATI, and to fields. Biol Cybern 27: 77– 87, 1977.
Baird JC, Taube JS, Peterson DV. Statistical and information properties of
minimize error, both should be equally strong. head direction cells. Percept Psychophys 63: 1026 –1037, 2001.
Bassett JP, Tullman ML, Taube JS. Lesions of the tegmentomammillary
circuit in the head direction system disrupt the head direction signal in the
Periodic input anterior thalamus. J Neurosci 27(28): 7564 –7577, 2007.
Bassett JP, Taube JS. Neural correlates for angular head velocity in the rat
The effect of adaptation and rebound for more general profiles is dorsal tegmental nucleus. J Neurosci 21: 5740 –5751, 2001.
complicated. However, in the special case that A ⫽ R and ␶ ' ␶A ⫽ Bassett JP, Zugaro MB, Muir GM, Golob EJ, Muller RU, Taube JS.
␶R, the combined effect of adaptation and rebound can be described as Passive movements of the head do not abolish anticipatory firing properties
a high-pass linear filter. This filter kernel is of head direction cells. J Neurophysiol 93: 1304 –1316, 2005.

J Neurophysiol • VOL 98 • OCTOBER 2007 • [Link]

1896 VAN DER MEER, KNIERIM, YOGANARASIMHA, WOOD, AND VAN ROSSUM

Blair HT, Cho J, Sharp PE. Role of the lateral mammillary nucleus in the rat McNaughton BL, Barnes CA, Gerrard JL, Gothard K, Jung MW,
head direction circuit: a combined single unit recording and lesion study. Knierim JJ, Kudrimoti H, Qin Y, Skaggs WE, Suster M, Weaver KL.
Neuron 21: 1387–1397, 1998. Deciphering the hippocampal polyglot: the hippocampus as a path integra-
Blair HT, Lipscomb BW, Sharp PE. Anticipatory time intervals of head- tion system. J Exp Biol 199: 173–185, 1996.
direction cells in the anterior thalamus of the rat: implications for path McNaughton BL, Battaglia FP, Jensen O, Moser EI, Moser M-B. Path
integration in the head-direction circuit. J Neurophysiol 78: 145–159, 1997. integration and the neural basis of the “cognitive map.” Nat Rev Neurosci 7:
Blair HT, Sharp PE. Anticipatory head direction signals in anterior thalamus: 663– 678, 2006.
evidence for a thalamocortical circuit that integrates angular head motion to McNaughton BL, Chen LL, Markus EJ. “Dead reckoning,” landmark
compute head direction. J Neurosci 15: 6260 – 6270, 1995. learning, and the sense of direction: a neurophysiological and computational
Blair HT, Sharp PE. Visual and vestibular influences on head-direction cells hypothesis. J Cogn Neurosci 3: 190 –202, 1991.
in the anterior thalamus of the rat. Behav Neurosci 110: 643– 660, 1996. Melvill Jones GM, Milsum JH. Frequency-response analysis of central
Boucheny C, Brunel N, Arleo A. A continuous attractor network model vestibular unit activity resulting from rotational stimulation of the semicir-
without recurrent excitation: maintenance and integration in the head direc- cular canals. J Physiol 219: 191–215, 1971.
tion cell system. J Comput Neurosci 18: 205–227, 2005. Mizumori SJY, Williams JD. Directionally selective mnemonic properties of
Brown JE, Card JP, Yates BJ. Polysynaptic pathways from the vestibular neurons in the lateral dorsal nucleus of the thalamus of rats. J Neurosci 13:
nuclei to the lateral mammillary nucleus of the rat: substrates for vestibular 4015– 4028, 1993.
input to head direction cells. Exp Brain Res 161: 47– 61, 2005. Muir GM, Taube JS. The neural correlates of navigation: do head direction
and place cells guide spatial behavior? Behav Cognit Neurosci Rev 1:
Brown JE, Yates BJ, Taube JS. Does the vestibular system contribute to head
297–317, 2002.
direction cell activity in the rat? Physiol Behav 77: 743–748, 2002.
Packard MG, McGaugh JL. Inactivation of hippocampus or caudate nucleus
Chen LL, Lin LH, Barnes CA, McNaughton BL. Head-direction cells in the with lidocaine differentially affects expression of place and response learn-
rat posterior cortex. II. Contributions of visual and ideothetic information to ing. Neurobiol Learn Mem 65: 65–72, 1996.
the directional firing. Exp Brain Res 101: 24 –34, 1994. Puccini G, Sanchez-Vives M, Compte A. Integrated mechanisms of antici-
Compte A, Brunel N, Goldman-Rakic PS, Wang X-J. Synaptic mechanisms pation and rate-of-change computations in cortical circuits. PLoS Comput
and network dynamics underlying spatial working memory in a cortical

Downloaded from [Link] on May 9, 2008

Biol 3: e82, 2007.
network model. Cereb Cortex 10: 910 –923, 2000. Ranck JB Jr. Head-direction cells in the deep cell layers of dorsal presubic-
Dudchenko PA, Muir GM, Frohardt RJ, Taube JS. What does the head ulum in freely moving rats. Soc Neurosci Abstr 10: 599, 1984.
direction system actually do? In: Head Direction Cells and the Neural Redish AD. Beyond the Cognitive Map: From Place Cells to Episodic
Mechanisms of Spatial Orientation, edited by Wiener S, Taube J. Cam- Memory. Cambridge, MA: MIT Press, 1999.
bridge, MA: MIT Press, 2005, p. 221–246. Redish AD, Elga AN, Touretzky DS. A coupled attractor model of the rodent
Etienne AS, Boulens V, Maurer R, Rowe T, Siegrist C. A brief view of head direction system. Network Comput Neural Syst 7: 671– 685, 1996.
known landmarks reorientates path integration in hamsters. Naturwissen- Séguinot V, Maurer R, Etienne AS. Dead reckoning in a small mammal: the
schaften 87: 494 – 498, 2000. evaluation of distance. J Comp Physiol [A] 173: 103–113, 1993.
Fernandez C, Goldberg JM. Physiology of peripheral neurons innervating Sekirnjak C, du Lac S. Intrinsic firing dynamics of vestibular nucleus
semicircular canals of the squirrel monkey. II. Response to sinusoidal neurons. J Neurosci 22: 2083–2095, 2002.
stimulation and dynamics of peripheral vestibular system. J Neurophysiol Sharp PE, Blair HT, Cho J. The anatomical and computational basis of the
34: 661– 675, 1971. rat head-direction cell signal. Trends Neurosci 24: 289 –295, 2001a.
Fuchs AF, Kimm J. Unit activity in vestibular nucleus of the alert monkey Sharp PE, Tinkelman A, Cho J. Angular velocity and head-direction signals
during horizontal angular acceleration and eye movement. J Neurophysiol recorded from the dorsal tegmental nucleus of Gudden in the rat: implica-
38: 1140 –1161, 1975. tions for path integration in the head direction cell circuit. Behav Neurosci
Gallistel CR. The Organization of Learning. Cambridge, MA: Bradford/MIT 115: 571–588, 2001b.
Press, 1990. Shimazu H, Precht W. Inhibition of central vestibular neurons from the
Georgopoulos AP, Kalaska JF, Caminiti R, Massey JT. On the relations contralateral labyrinth and its mediating pathway. J Neurophysiol 29:
between the direction of two-dimensional arm movements and cell dis- 467– 492, 1966.
charge in primate motor cortex. J Neurosci 2: 1527–1537, 1982. Skaggs WE, Knierim JJ, Kudrimoti HS, McNaughton BL. A model of the
Golob EJ, Taube JS. Head direction cells and episodic spatial information neural basis of the rat’s sense of direction. Adv Neural Inf Process Syst 7:
in rats without a hippocampus. Proc Natl Acad Sci USA 94: 7645–7650. 173–180, 1995.
1997. Song P, Wang X-J. Angular path integration by moving “hill of activity”: a
Gonzalo-Ruiz A, Alonso A, Sanz JM, Llinas RR. Afferent projections to the spiking neuron model without recurrent excitation of the head-direction
mammillary complex of the rat with special reference to those from system. J Neurosci 25: 1002–1014, 2005.
surrounding hypothalamic regions. J Comp Neurol 321: 277–299, 1992. Stackman RW, Taube JS. Firing properties of head direction cells in the rat
Goodridge JP, Taube JS. Preferential use of the landmark navigational anterior thalamic nucleus: dependence on vestibular input. J Neurosci 17:
system by head direction cells in rats. Behav Neurosci 109: 49 – 61, 1995. 4349 – 4958, 1997.
Goodridge JP, Taube JS. Interaction between the postsubiculum and anterior Stackman RW, Taube JS. Firing properties of rat lateral mammillary single
thalamus in the generation of head direction cell activity. J Neurosci 17: units: head direction, head pitch, and angular head velocity. J Neurosci 18:
9315–9330, 1997. 9020 –9037, 1998.
Goodridge JS, Touretzky DS. Modeling attractor deformation in the rodent Stringer SM, Trappenberg TP, Rolls ET, Araujo IET. Self-organising
head-direction system. J Neurophysiol 83: 3402–3410, 2000. continuous attractor networks and path integration: one-dimensional models
Hayakawa T, Zyo K. Afferent connections of Gudden’s tegmental nuclei in of head direction cells. Network Comput Neural Syst 13: 217–242, 2002.
the rabbit. J Comp Neurol 235: 169 –181, 1985. Taube JS. Head direction cells recorded in the anterior thalamic nuclei of
Johnson A, Seeland K, Redish AD. Reconstruction of the postsubiculum head freely moving rats. J Neurosci 15: 70 – 86, 1995.
direction signal from neural ensembles. Hippocampus 15: 86–96, 2005. Taube JS. Head direction cells and the neurophysiological basis for a sense of
Kaufman GD, Shinder ME, Perachio AA. Convergent properties of vestib- direction. Prog Neurobiol 55: 225–256, 1998.
ular-related brain stem neurons in the gerbil. J Neurophysiol 83: 1958 –1971, Taube JS, Muller RU. Comparisons of head direction cell activity in the
2000. postsubiculum and anterior thalamus of freely moving rats. Hippocampus 8:
Knierim JJ, Kudrimoti HS, McNaughton BL. Interactions between idio- 87–108, 1998.
thetic cues and external landmarks in the control of place cells and head Taube JS, Muller RU, Ranck JB Jr. Head-direction cells recorded from the
direction cells. J Neurophysiol 80: 425– 446, 1998. postsubiculum in freely moving rats. I. Description and quantitative analy-
Liu R, Chang L, Wickern G. The dorsal tegmental nucleus: an axoplasmic sis. J Neurosci 10: 420 – 435, 1990a.
transport study. Brain Res 310: 123–132, 1984. Taube JS, Muller RU, Ranck JB Jr. Head-direction cells recorded from the
Markham CH, Curthoys IS, Yagi T. The influence of the contralateral postsubiculum in freely moving rats. II. Effects of environmental manipu-
labyrinth on static and dynamic properties of brain stem vestibular neurons lations. J Neurosci 10: 436 – 447, 1990b.
in the cat, guinea pig and rat. In: Vestibular Mechanisms in Health and Tolman EC, Ritchie BF, Kalish D. Studies in spatial learning. II. Place
Disease, edited by Hood J. New York: Academic, 1978, p. 86 –94. learning versus response learning. J Exp Psychol 36: 221–229, 1946.

J Neurophysiol • VOL 98 • OCTOBER 2007 • [Link]

 HD ANTICIPATION MODEL 1897

Trappenberg TP. Fundamentals of Computational Neuroscience. Oxford, rotations: comparison with hippocampal place cells. J Neurosci 26: 622–
UK: Oxford Univ. Press, 2002. 631, 2006.
Vidal P-P, Sans A. Vestibular system. In: The Rat Nervous System (3rd ed.), Yu X, Yoganarasimha D, Knierim JJ. Backward shift of head direction
edited by Paxinos G. New York: Academic, 2004, p. 965–996. tuning curves of the anterior thalamus: comparison with ca1 place fields.
Wiener SI, Arleo A, Déjean C, Boucheny C, Khamassi M, Zugaro MB. Neuron 52: 717–729, 2006.
Optic field flow signals update the activity of head direction cells in the rat Zhang K. Representation of spatial orientation by the intrinsic dynamics of the
anterodorsal thalamus. Soc Neurosci Abstr 209.2, 2004. head-direction cell ensemble: a theory. J Neurosci 16: 2112–2126, 1996.
Wiener SI, Taube JS. Head Direction Cells and the Neural Mechanisms of
Spatial Orientation. Cambridge, MA: MIT Press, 2005. Zugaro MB, Arleo A, Berthoz A, Wiener SI. Rapid spatial reorientation and
Xie W, Hahnloser RHR, Seung HS. Double ring model of the head-direction head direction cells. J Neurosci 23: 3478 –3482, 2003.
system. Phys Rev E 66: 41902– 41910, 2002. Zugaro MB, Tabuchi E, Fouquier C, Berthoz A, Wiener SI. Active
Yoganarasimha D, Yu X, Knierim JJ. Head direction cell representations locomotion increases peak firing rates of anterodorsal thalamic head direc-
maintain internal coherence during conflicting proximal and distal cue tion cells. J Neurophysiol 86: 692–702, 2001.

Downloaded from [Link] on May 9, 2008

J Neurophysiol • VOL 98 • OCTOBER 2007 • [Link]