Integrating With Neurons: Ann. Rev. Neurosci. 1989. 12,' 33 - 45
Integrating With Neurons: Ann. Rev. Neurosci. 1989. 12,' 33 - 45
REVIEWS Further
Quick links to online content
Ann. Rev. Neurosci. 1989. 12,' 33--45
Copyright © 1989 by Annual Reviews Inc. All rights reserved
D. A. Robinson
Annu. Rev. Neurosci. 1989.12:33-45. Downloaded from www.annualreviews.org
INTRODUCTION
yet) = L x(r) dr 1.
in the steady state. On the other hand, integrators have low gains at high
frequencies, helping to prevent oscillations. If a control system does not
contain an integrator naturally (such as a motor), the design engineer
will probably add one, if not two, to the system to get the desired perfor
mance.
Consequently, when engineers became interested in biological control
systems, they took it for granted that integrators were everywhere-how
else could all these control systems possibly work? The oculomotor system
Access provided by 2806:106e:15:754d:d90d:36ad:c2dc:bd1f on 05/25/20. For personal use only.
offers an especially clear example: The retina senses the error between the
eye (fovea) and the target, and the system turns the eye until the error is
Annu. Rev. Neurosci. 1989.12:33-45. Downloaded from www.annualreviews.org
zero--a simple negative feedback scheme (e.g. Young & Stark 1963).
Moreover, when the goal is reached, a constant eye deviation (output) is
maintained while the error (input) is zero. But that is just what an integrator
does-it holds signals in the absence of new information. Indeed, the only
way its output can be constant is if the input error is zero-the desired
condition. So obviously the oculomotor system had an integrator and to
the bioengineer this idea was so obvious as to be trivial.
To the neurophysiologist a neural integrator seemed exotic, but, so long
as it only appeared in top-down, black-box models, it could be relegated
to "higher centers" and ignored. This was not, however, the case for the
vestibulo-ocular reflex. By the early 1960s, it was clear that the signal from
the semicircular canals, coded in the rate modulation of the primary
afferents, was head velocity. The discharge rate of motoneurons, on the
other hand, largely determined eye position. If a constant head velocity is
to make the eyes move at a constant velocity, the motoneurons must
respond to the time integral of the canal signal. This integrator was not
hidden under a feedback loop-it was the major signal-processing element
in a short, well-defined, pontine reflex. It could not be ignored without
also ignoring the main function of the reflex.
Still, it was not until 1968 that this obvious observation first appeared,
however briefly, in print (Robinson 1968). Soon thereafter neuro
physiological evidence appeared. Cohen & Komatsuzaki ( 1972) found that
electrical stimulation of the reticular formation caused the eyes of monkeys
to move at a constant velocity-the time integral of the step of excitation.
We know now that they were stimulating an input to the integrator (from
saccadic burst neurons), not the integrator itself, but no matter; integration
was clearly occurring. It was also confirmed that motoneurons were
responsible for determining the position of the eye (Robinson 1970), and
by recording from them during sinusoidal rotations of monkeys we
demonstrated the requisite 90 deg phase lag between vestibular and oculo
motor motoneurons created by the integrator (Skavenski & Robinson
1973).
INTEGRATING WITH NEURONS 35
Although the integrator's existence was not open to question simply on
theoretical grounds, these findings lent a sort of respectability to the idea
and helped in making clear the integrator's essential role. The concept
was readily accepted in neuro-ophthalmology; after all, it is the neural
integrator that generates the slow phases of nystagmus and holds the eye
eccentrically after a saccade. Disorders of these basic operations could
now be attributed to the integrator.
Access provided by 2806:106e:15:754d:d90d:36ad:c2dc:bd1f on 05/25/20. For personal use only.
Figure 1 shows the signal processing involved in the reflex. On the right, the
canal produces, in the frequency range of physiological head movements, a
signal proportional to head velocity, H(t), coded in the discharge-rate
modulation. Rv" of primary vestibular afferents. The background rate
(90 spikesjsec) and sensitivity [0.4 (spikesjsec)j(degjsec)] are taken from
Fernandez & Gol dberg (197 1). On the left is shown the well-established
relationship between the modulation of motoneuron discharge rate, llRm
flocculus
i'"_ �
y vermis
E�-_�_ ��'
mn pursuit
R VI =90+0AH
/ '/
I y; I"'n�
I
\ _J_�d�)
.
H )
NI
conol
Figure 1 The final common integrator. On the right the canals transduce head velocity, H,
and report it, coded as the modulation of the discharge rate, R", of primary vestibular
afferents to the vestibular nucleus, vn. This signal becomes an eye velocity command for
vestibular movements, Ev, which is sent directly to the motoneurons, mn, and to the neural
integrator,NI, to provide the needed position signal E. These signals provide those needed
by the motoneurons modulating by /':,.Rm. The pursuit and saccadic signals also arrive as eye
velocity commands, Ep and E" from the cerebellum (flocculus, vermis) and paramedian
pontine reticular formation (pprf), respectively, and are also sent to the motoneurons and
the neural integrator. The latter is contained in the nucleus prepositus hypoglossi and the
vestibular nuclei (dashed lines).
36 ROBINSON
that the neurons in the vestibular nucleus project directly to the moto
neurons to form the well-known three-neuron arc.
Annu. Rev. Neurosci. 1989.12:33-45. Downloaded from www.annualreviews.org
For a long time the neural integrator was thought to lie in the paramedian
pontine reticular formation because lesions there affected eye movements
profoundly and the reticular formation seemed a good place for anything
mysterious. This idea was finally disproved when Renn and his colleagues
made neurotoxin lesions there. Ipsilateral saccades were abolished but not
gaze holding (Renn et aI1984). The cerebellum is important in minimizing
Access provided by 2806:106e:15:754d:d90d:36ad:c2dc:bd1f on 05/25/20. For personal use only.
sec. This might tempt one to put the integrator entirely in the cerebellum
(Carpenter 1972) except that 1.3 sec is not negligible. During vestibular
nystagmus, the integrator need only integrate well from one quick phase
to the next (roughly 0.3 sec), and even after total cerebellectomy only close
examination will detcct such lcakiness in nystagmus recordings. Moreover,
during ice-water stimulation of one ear, simulating a vestibular lesion, Tn
in humans is deliberately decreased to about 2.4 sec within 80 sec of the
onset of the inappropriate nystagmus (Robinson et al 1984). The cere
bellum may well be responsible for parametric adjustments with this time
scale. It is shown below that lesions of the vestibulo-prepositus complex
abolish all integrator action, so the current thinking is that the integrator
is basically in this complex, but the cerebellum has a powerful influence in
adjusting its time constant.
Studies of cells in the vestibular nuclei of alert monkeys discovered that
a large proportion of cells in the superior and rostral-medial subdivisions
carried, among others, the eye position signal E. Many cells are purely
oculomotor in that their activity reflects eye movements whether or not of
vestibular origin. Thus, significant subdivisions of these nuclei form an
eye-movement nucleus. This led Tomlinson & Robinson (1984) to propose
this region as the site of the neural integrator. Meanwhile, the nucleus
prepositus hypoglossi, right next door, had also been suggested because
its cells also carried the eye-position signal and projected directly to moto
neurons. Finally, Cheron et al (1986) showed integrator failure after electro
lytic lesions of either region in the cat, and Cannon & Robinson (1987)
showed in the monkey total loss of the neural integrator after bilateral
neurotoxin lesions of both the prepositus and medial vestibular nuclei.
The latter study showed that, as one would predict from Figure 1,
without the integrator the eye velocity commands would pass directly to
the motoneurons and produce an eye position that was proportional to
desired eye velocity. For a step of head velocity, for example, thc step of
Ev, in the absence of the ramp normally generated by the integrator, simply
causes a step change in eye position without the usual slow phases of
38 ROBINSON
the outputs, x let) and xlt), respond with a time constant of r/ ( l + w) or,
since w is close to 1.0, about 2.5 msec. Thus, the background rates are
Annu. Rev. Neurosci. 1989.12:33-45. Downloaded from www.annualreviews.org
time constant of ,/ ( 1- w). If w is very close to 1.0, this time constant can
be very large, and integration takes place. It is remarkable that lateral
inhibition, a work horse of neural network modelers, solves this problem
as well.
The second problem is that to increase the time constant from 5 msec
to 20 sec, the value of w must be 0.99975. Of course, if the membrane time
constant of these particular neurons was 50 instead of 5 msec, the value
of w becomes 0.9975, but this is still hardly robust. On the other hand, no
one supposes that two neurons are adequate. Cannon et al (1983) examined
a ring of 32 neurons connected as in Figure2B. Figure2C shows that the
Bode diagram (log gain vs log frequency, OJ, if the input were a sinusoid)
depends on the spatial frequency, P, of the inputs. If all the cells received
the same signal (e.g. the background rate), its spatial frequency would be
zero and the network would act like a wide-band, low-gain system with a
time constant of2.5 msec (front edge in Figure2C). If the inputs from left
and right canals (or pursuit cells or saccadic burst neurons) are interleaved
as shown, one has the highest spatial frequency (n/2 or one half cycle/
neuron), and the system has the Bode diagram of an integrator with a time
constant of 20 sec (back edge in Figure 2C).
We were able to show that adding spatial noise to such a model (changing
synaptic strengths randomly throughout) did little to perturb its integrating
behavior. Nothing depended heavily on only a few synapses. Killing one
cell (out of 32, 3%) caused the impulse response (a saccade) to drift back
quickly by about 15%, because the dead neuron broke local feedback
loops with nearby cells, but then to recover and drift toward zero at a rate
appropriate to a 20 sec time constant. This result indicated that the major
property of the network did not depend critically on any one parameter
in the model. The model was robust.
Integration still requires a critical adjustment in a global sense. If Xes, P)
40 ROBINSON
X X
I 2
t::= ,
� t
X
I
X
2
C
A
W
Access provided by 2806:106e:15:754d:d90d:36ad:c2dc:bd1f on 05/25/20. For personal use only.
I I
Annu. Rev. Neurosci. 1989.12:33-45. Downloaded from www.annualreviews.org
u, u, u. u.
� �
Figure 2 A model of the neural integrator. A: Lateral inhibition. Each cell excites itself
(positive feedback) by self-disinhibition through its neighbor. A common input, UI and U2,
is not integrated but is simply repeated by the outputs x I and x2• A differential input, on the
other hand, such as flu, is integrated as shown. B: A fuller model with 32 neurons. The
spatial variablek runs from I to 32. The neuron at ko inhibits itself and neighboring cells
with the strength w(k-ko) shown by the double Gaussian curve at the top. The inputs enter
from push-pull sources and alternate on the neurons to achieve the highest spatial frequency:
nl2 or one half cycle per neuron. C: The Bode diagram of the gain XI U of this network with
the spatial frequency of the input P as a third dimension. As P approaches n12, the system
behaves like an integrator with a time constant T of 20 sec.
INTEGRATING WITH NEURONS 41
and U(s,P) are the transforms of the outputs and inputs, where s is the
Laplace transform frequency in the temporal frequency domain and P is
the Fourier transform frequency in the spatial frequency domain, then
X(s,P)
2.
U(s, P) S7: + [1- W(P)]
where W(P) is the Fourier transform of the spatial function of the pattern
Access provided by 2806:106e:15:754d:d90d:36ad:c2dc:bd1f on 05/25/20. For personal use only.
trying to make midline lesions to interrupt these fibers, have met with
peculiar results. In only one monkey so far, electrolytic lesions in the
midline, in the region 0 to 4 mm caudal to the abducens nuclei and 4 mm
deep, did decrease the integrator time constant to 0.6 sec after each lesion,
but recovery was significant in just a few hours and more so after two
days. Two nagging questions arise: What is the anatomical course of these
crossing fibers-did we interrupt them? What pathways and mechanisms
mediate this incredible capacity for repeated recuperation?
Equally important, new models should be proposed by theoreticians to
provoke further experimental tests. The operation we are considering is
not trivial.
INTEGRATORS ELSEWHERE
over, the canal signal rides on a high, stable, resting background rate so
that modulation both up and down, in push-pull, can occur, thereby
Annu. Rev. Neurosci. 1989.12:33-45. Downloaded from www.annualreviews.org
Literature Cited
ander's law: Its behavior and origin in and paraflocculus on eye movement in pri
the human vestibulo-ocular reflex. Ann. mate. J. Neurophysiol. 46: 878-99
Neurol. 16: 714-22