Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

IEEE T R A N S A C T I O N S ON B I O M E D I C A L E N G I N E t R l N G . VOL 36.

N O I , J A N U A R Y 1989

Dynamical Neural Network Organization of the


Visual Pursuit Svstem J

L. KELLER, AND WILLIAM F. CRANDALL

Abstract-The central nervous system is a parallel dynamical system use ideas from control systems engineering. For example,
which connects sensory input with motor output for the performance Cannon and Robinson [4] use state feedback and pole-
of visual tracking. This paper applies elementary control system tools
to extend dynamical neural network models to the visual smooth pur-
placement to form an integrator from biologically plau-
suit system. Observed eye position responses to target motions and sible elements. Closer to the present paper is an interest-
characteristics of the plant (eye muscles and orbital mechanics) place ing simulation by Barto et al. [5] of an inverted pendulum
dynamical constraints on the interposed neural network controller. In controller with parameters “learned” by a neural net-
the process of constructing a model for the controller, we show two work. Sompolinsky and Kanter [6] describe a neural net-
previous pursuit system models, using efference copy and feedforward
compensation, are equivalent from an input-output standpoint. We in-
work which is capable of recalling temporal patterns and
troduce a controller model possessing a potentially highly parallel im- suggest this may be usefui in understanding rhythmical
plementation and offer an example with supporting neural firing rate patterns in biological motor systems.
data. Changes in time delays or other system dynamics are expected to The visual tracking system is attractive because it ap-
lead to compensatory adaptive changes in the controller. A scheme to pears to consist of a simple sensory input to motor output
noninvasively simulate such changes in system dynamics was devel-
oped. Actual physiologic data of adaptive responses to increased time
loop. To achieve good tracking behavior, the brain must
delay is presented as an example of the utility of this parallel control- serve as a controller which compensates for time delays
ler. Compensatory changes in our parallel controller model are easily and dynamics and is capable of adaptively responding to
predicted. These results suggest a productive interaction between changes in time delay or dynamics. From this perspec-
neural network modeling, neurophysiology, and control systems engi- tive, the brain is a dynamical, highly parallel, adaptive
neering.
control system. This paper develops a dynamical neural
network which helps to unify two apparently disparate
I. INTRODUCTION models of smooth pursuit and clarify dynamical element
organization. Realization issues are addressed, physio-
T HE FIELD of neural networks is characterized by ex-
plorations of properties of many simple dynamic sys-
tems interconnected to collectively form a system which
logic signal correlates suggested, and a framework offered
to understand how adaptive changes might be imple-
does computation. One common form can be considered mented to compensate for eye muscle or neural dynamics
to be a function from a set of input vectors to a set of changes.
output vectors. These networks serve as associative mem-
ories, pattern recognizers, and optimization schemes. A . Experimental Methods
Good examples in this area include Kohonen and Oja [ 11, Experimental data and models developed in this report
Anderson [ 2 ] , and Hopfield and Tank [3]. Dynamical is- are taken from eye movement recordings in alert behaving
sues in such systems concern convergence behavior in monkeys. Human pursuit eye movement responses are
“training” sessions or convergence in pattern completion similar, but slower. This paper concentrates on the pursuit
or optimization solutions. response to ramp target motions where the visual target
Less common are neural net applications where dynam- begins to move at a constant but unpredictable speed, at
ics are central to the goal of the network. Here the net- a random epoch, and in an unpredictable direction. Pur-
work is considered to be a functional from a set of inputs suit responses to predictable target motion probably in-
(over time) to a set of outputs. Such applications often volve a predictive neural network that is not considered
here.
Manuscript received February 8, 1988: revised July I , 1988. This work The pursuit eye movements studied in this report were
was supported by Grants National Eye Institute EY0S913, EY06860.
EY06883, and EY06860, the Rachael Atkinson Fellowship, and the Smith-
generated by visual target motion produced by an oscil-
Kettlewell Eye Research Foundation. loscope projector system [7]. Eye position was measured
D. C . Deno is with the Department of Electrical Engineering and Com- by the implanted induction eye coil technique [8].
puter Science, University of Berkeley, CA 94720.
E. L. Keller is with the Department of Electrical Engineering and Com- Changes in plant dynamics were electronically simulated
puter Science, University of California, Berkeley, CA, and the Smith-Ket- by modifying target position based on eye position sig-
tlewell Eye Research Foundation. San Francisco, CA 941 IS. nals. Data were collected by digitizing the eye and target
W . F. Crandall is with the Smith-Kettlewell Eye Research Foundation,
San Francisco. CA 941 IS.
position and velocity signals at 0.5 or 1 kHz sampling
IEEE Log Number 8824419. rates. Real-time experiment control and data acquisition

0018-9294/89/0100-0085$01.OO 0 1989 IEEE


86 I E E E T R A N S A C T I O N S O N B I O M E D I C A L E N G I N E E R I N G , VOL. 36, N O . I. J A N U A R Y 1989

respect to target velocity, the subsequent damped oscil-


lation of eye velocity about true target speed, and the final
close approximation of steady-state eye speed to target
speed. All these features are similar to those previously
reported in man [12], [13] and monkey [14], [7].

11. THEORETICAL DEVELOPMENT


Time (msec) from Target Motion A . Oculomotor Physiology
Fig. I . A representative smooth pursuit system response. Velocity is in
degree/second. A 2 O o / s step in target velocity is shown as the dashed to this system is the strut-
lines. The eye velocity recording is the jagged curve and the blank space turd subtraction of target and eye position which occurs
in this recording shows where a saccade was removed. The solid smooth on the retina as shown in ~ i qa). ~ The. current dogma is
curve shows the best nonlinear least-squares fit to the underlying data.
Even within a set of trials in one animal there is Some variability in exact that this visual error information is the primary input into
delay, damping, and oscillation frequency. Since point-by-point aver- the pursuit system [ 141 and that neither target nor eye po-
aging in the face of this variability can obscure significant features we sition information are directly available to the
use analyses of individual trials to yield parameterizations from which a
. I . .
Drototvpe response is derived from the median or mean parameter set. system. From photoreceptors information then
Our model simulations are based on similarly generated prototype re- travels to visual cortex and then to temporal cortical re-
sponses. gions believed specialized for motion analysis [ 151 where
cells respond preferentially to certain target speeds and
were accomplished by a set of laboratory computers (Ap- directions and are less influenced by retinotopic location.
ple 11, IBM-AT, and a PDP-11). Linear system manipu- These temporal cortex cells respond to target motion with
lation was aided by Program CC [9] and nonlinear nu- delays on the order of 80 ms. This velocity error infor-
merical simulation performed with Simnon [ IO]. mation is believed to be a major drive behind the smooth
pursuit system. Together with target position information
B. Smooth Pursuit Major Attributes and higher cortical inputs such as volition, the velocity
The visual tracking system provides eye movements information then passes to the pontine region of the brain
which keep images of moving objects near the fovea where stem and then to the cerebellum where it interacts with
visual acuity is highest. The eyes move in response to the signals helping to coordinate information that now
target motion in a feedback-control system which seeks specifies motor commands [ 141. The information is then
to reduce fixation error. Eye movements take several reintegrated in the brain stem and sent to oculomotor nu-
forms: smooth pursuit (sustained, low velocity motions clei where the information travels through the cranial
which help match eye and target velocity), saccades (ep- nerves to the eye muscles which rotate the eye.
isodic, high velocity movements to reduce position error),
and vergence (disconjugate movements which align im- B. Development of a Mathematical Model
ages from near-field objects on the two retinas). The From the conceptual information flow diagram of Fig.
smooth pursuit system has proven to be a fruitful subject 2(a) we incorporate some linear systems dynamical no-
of study from both neurophysiology and engineering per- tation in Fig. 2(b). Cortical motion processing, convert-
spectives. ing position information into velocity information, is rep-
When tracking is elicited by a sudden target motion, the resented by a differentiator ( s in Laplace notation). The
usual response is to attempt to reduce this error by eye loop delay has been lumped into a single element e-sT.
movements consisting of both smooth pursuit and sac- The transition from sensory information into motor infor-
cadic components. Smooth pursuit eye movements with- mation is begun in the neural controller element labeled
out corrective saccades can be induced by superimposing Gc. This scheme also includes the generally accepted no-
a step in target position away from the fovea in the direc- tion that neural circuits which immediately precede ocu-
tion opposite to target motion (the Rashbass step-ramp lomotor neurons provide both direct (velocity, 7 )and in-
paradigm [ I I]). If the step size is chosen so that the eye tegrated (position, 1 / s ) signals to these neurons. The
is just beginning to move as the target crosses the fovea, “neural integrator” is assumed to have a time constant
then the saccadic portion of the pursuit movement will be much larger than 1-2 s and can thus be approximated by
suppressed. an ideal integrator 1 /s. From data collected in vergence
Fig. 1 shows a typical response in one monkey to the studies [16] it appears that a step change in motor unit
step-ramp paradigm when the visual target velocity was firing rate results in eye position changes well modeled as
2Oo/s to the left. A major feature of the response is the a two-pole plant with time constants of 0.2 and 0.015 s.
delay from the onset of target motion to the apparent ini- Higher order lumped parameter models of the eye, mus-
tiation of the response. This delay is typically 100 ms but cles, and connective tissue [17] differ little in predicted
can be as short as 80 ms or as long as 150 ms in some responses. Nonlinearities in the plant due to saturation are
animals. Other features of the pursuit responses are the unlikely in smooth pursuit where velocities are an order
rapid initial eye acceleration immediately following the of magnitude less than during saccades.
period of the delay, the overshoot of eye velocity with Fig. 2(c) shows the (parallel) premotor path effectively
DEN0 el u l . : NEURAL NETWORK ORGANIZATION OF VISUAL PURSUIT SYSTEM 87

Fig. 2. Basic structure of the smooth pursuit system. (a) Information flow
diagram. Retinal position information is the difference between target
and eye position. Representation of this information is unspecified arid c- "G ,)
nonlinear dynamics are permitted. Abbreviations: cerebel. (cerebellum);
integr (integrator); propor. (direct proportiona' pathway). (b) Mathe-
matical model using Laplace notation. Delay has been lumped into a
single element located where the majority of delay exists. The s denotes
differentiation to obtain velocity information. The controller Gc is un-
specified and possibly nonlinear. The plant GP is replaced with a linear
systems model. Abbreviations: T(target position); E (eye position); RPE
(retinal position error). (c) Equivalent model. Velocity can be considered
to be the input and output of the smooth pursuit system. The premotor
integral and direct paths create a zero which may cancel the plant's slow
pole 7 ,since neural signals preceding this point seem to code for desired
eye velocity. (d) Final reduced model. Models of the pursuit system share Fig. 3. Various controller models. (a) Simple proportional gain. (b) Rob-
this basic arrangement but differ in realizations for Gc and G p . The con- inson et al. [I31 (linearized). (c) Lisberger er al. [21] (linearized). (d)
troller has input node A (delayed retinal velocity error, RVE) and output Alternative cancellation scheme. A rearrangement of cancellation
node B (eye velocity command). schemes to emphasize three main components: G&! which embodies a
(inverse) model of the plant, e-'TMwhich is a model of total loop delay,
and GDa desired response element (e) Partial fraction re-representation
of D above. The straight forward compensation in D above can be dis-
canceling the slow dynamics of the plant, as is commonly guised as a parallel set of dynamical elements.
assumed [ 131. It also shows that the input and output sig-
nals can be considered to be target velocity T and eye
velocity E , consistent with conventional descriptions of to account for observed periods of steady-state tracking
the smooth pursuit system. Finally, this simplified con- where E 1 T[13].
ceptual scheme can be formatted as shown in Fig. 2(d) in The second class of models, illustrated in Fig. 3(b),
which velocity inputs and outputs are explicitly shown. may be called an efferent copy feedback model. A copy
Whether or not there is cancellation of the neural integra- of the neural command for desired eye velocity (node B )
tor by differentiation and cancellation of a plant pole by is passed through a central model of the plant dynamics
the premotor path, the resulting scheme may still be rep- and total loop delay e-"'GpM. The resulting output signal
resented as in Fig. 2(d). All models and simulations of summed with the delayed retinal velocity error signal
the system can start with this basic topology-they will (node A ) recreates a central replica of target velocity TR.
differ only in the dynamics (possibly nonlinear) assigned This is a valuable transformation and some oculomotor
to the controller Gc and the premotor circuit/plant Gp. neurophysiology supporting a reconstructed TR has re-
cently been described in the cerebellar vermis by Suzuki
C. Previous Pursuit System Models and Keller [20]. Taken together, Figs. 2(d) and 3(b) in-
Models of the pursuit system can be divided into three dicate that this arrangement creates an effectively open-
broad classes [Fig. 3(a)-(c)] based on the controller model loop system in which reconstructed target velocity di-
Gc. Initial efforts to model smooth pursuit tracking de- rectly generates eye velocity through a central processor
scribed the system as a simple proportional error feedback G;7. It is then clear that all the dynamic features observed
servo that attempted to reduce the retinal slip velocity of in E (Fig. 1) for step inputs in target velocity must be
the selected target image [18]. Fig. 3(a) illustrates a sim- generated by Gh. The recent model by Robinson et al.
ple velocity-servo in which Gc becomes a scalar gain fac- [13] uses this sort of efferent feedback. By judicious in-
tor K . This model fails on two counts. First, there is a clusion of nonlinear elements their model displays appli-
stability incompatibility in the requirement for both high cability over an extended range of target velocities.
gain and a significant delay [ 191. Second, the model fails The third class of model is exemplified by Lisberger et
88 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING. VOL. 36. NO. I , JANUARY 1989

al. [21] as shown in Fig. 3(c). It uses the retinal motion ical concepts. One question that arises from cancellation
signal as the input to a parallel network which produces models such as Fig. 3(b) and (d) is how to create internal
weighted combinations of retinal position, velocity, and model delays on the order of 100 ms? It seems awkward
acceleration errors (RPE, RVE, and RAE respectively). to string together 100 1 ms synaptic delay elements and
These signals are then combined and fed through a final switch some in or out to adapt to different delays. Alter-
‘‘velocity memory” integrator to produce the eye velocity natively, one may approximate delays by systems which
command. The necessity for the final integrator stems use a sort of mutual cancellation, e.g., Pad6 approxima-
from open-loop pursuit experiments of Lisberger et al. tions. Time delays realized in this manner make the tran-
[21]. If the visual feedback loop is electronically opened sition to our parallel neural network controller of Fig. 3(e)
by fixing the target’s position on the retina, the animal’s natural.
eyes continue to move for a few seconds. Since retinal
A . Controller Example
position and velocity error can be fixed at zero, some type
of velocity holding circuit is needed and can be modeled An example may help clarify our approach. A prototype
as an integrator. This model is attractive because the coef- smooth pursuit response (similar to Fig. 1) for tracking a
ficients of RPE, RVE, and RAE were directly estimated 10-20” /s targets was well fit by a linear, time-invariant,
by experimentally imposing controlled position, velocity, deterministic, continuous system with transfer function
or acceleration errors on the retinal input during open-
~-”G,(s) =
e-’.’”(5s +200)
loop tracking. This model does not seem to extend as
readily to a wide range of velocities as the efferent feed-
+ +
0 . 0 0 3 ~ ~0 . 2 1 5 ~ ~6s + 200’
back type [ 131, but it too can produce many features of A single-pole linear plant model
the monkey smooth pursuit response. 1
D. A Parallel Pursuit System Model Gp(S) = 0.015s +1
We introduce a variation on the controller schemes de- is assumed as motivated earlier and thus it is straight for-
scribed above which is shown in Fig. 3(d). For both con- ward to determine Gc from Fig. 3(d). Replacinge-”M with
ceptual simplicity and because signals which precede the a fifth-order Pade approximation leads to an 8th-order ra-
neural integrator seem similar to eye velocity, we precede tional polynomial expression for Gc whose partial frac-
the plant GPwith an inverse plant model GiL. For a linear tion expansion consists of five terms. Under the condition
plant this implicitly requires GP be stable and minimum where delays T = T, = 0.1 s, a parallel controller rep-
phase. At the input of this inverse plant is thus a central resentation [as in Fig. 3(e)] is
replica of eye velocity which, when passed through a
model of the loop delay e-FTM, can cancel retinal feedback G,(s) =
9.524
+ 8.659s + 713.5
+ 42.632
~

S (s - 2.715f
and thereby produce a central replica of the target’s ve-
locity. The element labeled Go denotes the desired dy-
+ + 695.6
.0436s
namics (possibly nonlinear) for the pursuit response.
As a conceptual tool, and for correspondence with (s + 22.59f + 93.762
neural recording data discussed later, we develop a par-
allel feedforward dynamical neural network representa-
+ 6.777s + 1402 - .00356
(s + 146.2f + 100.82 S + 39.57’
tion for this controller. Assuming Go is linear and choos-
ing rational polynomial approximations for GD and e-’TM, Because the first two terms dominate, the controller could
a partial fraction expansion of Gc leads to a network of consist of two parallel dynamical pathways.
the form shown in Fig. 3(e). This basic manipulation If T and T, were increased to 0.18 s, and the controller
shows, in the linear case at least, that the two classes of adaptively responded to preserve the same step response
models (parallel feedforward and delayed efferent feed- (except for an additional 80 ms delay) the controller rep-
back) can be shown to be equivalent. This creates a bridge resentation would be
between the Lisberger and Robinson perspectives. Fur-
thermore, issues of plasticity and adaptation are more eas-
5.405
G ~ ( s=) - + 12.90s + 220.1
ily addressed from the twin perspectives of Fig. 3(d) and
S (s - 3.326f + 29.752
(e). One can immediately predict what a corrective re- 0.2078s + 734.2
sponse to a change in system time delay, gain, or dynam- +
ics would entail (as done below). Finally, we are free to + 4.563)2 + 58.772
(s

+ 6.483s + 1107 - 0S.000000


select Go to match an arbitrary input-output pair, con- 12
trary to the claim by Robinson et al. [ 131.
( s + 97.93)2 + 79.102 + 40.00 ‘

111. NEURALNETSA N D NEUROPHYSIOLOGY Here, in addition to the first two terms, the third term
Models constructed from input-output responses are not plays a significant role. An adaptive change in Gc might
unique. Realizations are often selected based on simplic- consist of changing the handful of parameters involved in
ity and consistency with both current data and physiolog- these three terms.
D E N 0 er ul : NEURAL NETWORK ORGANIZATION OF VISUAL PURSUIT SYSTEM 89

B. Neurophysiologic Correlates
One possibility is that there are sets of brain cells whose
m
discharge envelopes approximate each of the five hypoth-
esized parallel signals in our expansion. There are, how-
I I
ever, many equivalent schemes where cells carry linear
combinations of this model's paths.
For the case where T = 0.1 s and two parallel controller 0 l5 1
paths predominate we next illustrate that neural firing rate
data can be consistent with the partial fraction expansion
above. The top panel of Fig. 4 shows the location of the
two significant outputs which together comprise the con-
troller. In the middle and lower panels of Fig. 4, the out-
puts of the first and second terms are paired with neural
discharge records actually recorded in the brainstem of a
monkey shown such a target motion. These cellular re-
sponses are examples of a variety of dynamic neural sig-
nals recorded with microelectrodes in the pontine nuclei
of monkeys (Keller and May, unpublished data). This re-
gion of the primate brain receives parallel inputs from a
wide variety of extrastriate cortical regions thought to be
involved in processing visual motion information. It has Time (msec) from Target Motion
outputs to regions of the cerebellum involved with smooth Fig. 4. Parallel controller diagram and outputs. The plant was assumed to
eye movements [22]. Thus, hypothetical responses car- be linear single pole, the desired response assumed to be a linear second-
ried in these two paths may have some basis in neuro- order system, and the model time delay replaced by a fifth-order Pad6
approximation. The resulting transfer function from A to E , after partial
physiology. fraction expansion, consisted of only two dominant terms or elements.
Without reference to the model, these neural responses Top: the two parallel controller elements G,, and Gcz are shown within
could be difficult to fit into a functional relationship with the pursuit system model. Middle: the G,, term was effectively an in-
tegrator and its output is shown by the smooth curve. It is matched with
the generation of pursuit. A direct interpretation of the neural firing rate data from single-unit recording during such a pursuit
neural response shown on the middle panel of Fig. 4 is as trial. Both records are synchronized on onset of target motion. Lower:
the integral of the retinal velocity slip signal. The func- the other term GC2was a second-order system with an output shown by
the smooth curve. Again, neural firing rate data from single-unit record-
tion of the cell depicted in the lower panel may be to pro- ing to show the existence of units with activity that can plausibly resem-
vide an early boost in eye velocity (at the cost of later ble the outputs of such simple parallel dynamical elements. All controller
oscillations in eye velocity about the true target velocity). outputs are to the same scale.
Note that this neural signal is not in phase with either eye
acceleration or retinal acceleration error, as reflected in
the expression for G,( s). In fact, in a highly parallel sys- flect sensitivity to an imperfect internal model, and the
tem each neuron engaged in control could possess a dif- long term adaptation to such a change would reflect a re-
ferent response pattern. This highlights one of the major calibration of this internal model.
differences of our parallel model from that of Lisberger et It is not possible to actually modify the plant's dynam-
al. [21] where only retinal velocity error and its deriva- ics or delay in normal human subjects or without invasive
tives and integrals make up the parallel paths. procedures on primates. The diagrams of Fig. 5 show an
approach to provide the retina with information consistent
IV. CALIBRATION
A N D PLASTICITY
with modified plant dynamics. To the extent that the ret-
ina constitutes the sole or major feedback pathway in pur-
The requirement for calibration to maintain accurate in- suit, this external manipulation must completely simulate
ternal correlates of dynamic elements is often pointed out the desired change to the plant.
in discussions of the vestibulo-ocular reflex (VOR) (e.g., Consider a modified plant whose dynamics are that of
[24]). The plasticity of neural gains in the VOR has been the original plant in series with a new element wch ad-
clearly demonstrated by the telescopic spectacle and sim- ditional dynamics, gain, and time delay Gp =
ilar paradigms. Adjustments in synaptic efficacy are con- GpGAePSTA [Fig. 5(a)]. The old plant is a special case
sidered likely mechanisms for gain alterations required to where GA(s) = 1 and TA = 0. Next, we move the
keep the essentially open-loop VOR accurate. GAe-sTAblock maintaining the same retinal signal and
The utility of compensators that employ (explicitly or overall input-output relation. As shown in Fig. 5(b), this
implicitly) models of the plant's delay and dynamics was results in an element that uses current and past eye posi-
discussed above. The models proposed did not include tion to modify the actual target position T (perc%ived by
descriptions of how such internal models are kept in ac- the subject) from the "planned" target position T :
curate form. Upon suddenly perturbing the system's delay
or dynamics, the immediate resulting response should re- T = f + ( 1 - GAe-"A)E.
90 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 36, NO. I , JANUARY 1989

E 40 1 n
:
V
e
1

r - - - - - - - - -I

E
GAe-"b
B
z
V
e
1
0
C
(b)
Y
i
Fig. 5. Target modification scheme to permit simulation of altered plant
dynamics. (a) Motivational diagram. Consider the old plant G , to be in
series with a new dynamical element GAe-'TAthereby forming the new E 60
plant Gp. (b) Equivalent diagram. Simulation of altered plant by elec-
tronic modification of target position utilizing an element with transfer :
function 1 - G,e-"7'. The term G , may incorporate dynamics (possibly V 40
nonlinear). e
1
0
F:

We have implemented digitally an element G,( s) = 1 - t.


Y
GAePSTA that operates at rates of 1-2 kHz for both vertical -200 0 200 400 600 800 1000 1200
and horizontal directions and allows selection of nearly Time (msec) from Target Motion
any time delay, gain, and single-pole location. Fig. 6 . Illustration of time delay adaptation. Velocity is in degree/second.
Some preliminary results for delay modification are Top: fairly oscillatory smooth pursuit response under control conditions
presented in Fig. 6. The top panel shows a fairly oscil- to a 20"/s target. The smooth curve is the result of a least-squares fit to
the data from -200 ms to 1200 ms. Middle: same monkey as above with
latory pursuit trial without saccades and under the control TAacutely set to 80 ms to simulate an extra 80 ms delay. The oscillation
condition GA = 1 and TA = 0. The smooth curve is a period has gone from nearly 300 ms above to about 600 ms. Lower: after
least-squares fit of a second order linear system step re- three days of three-hour training sessions with TA at 80 ms a typical
pursuit response now has an oscillation period of about 400 ms indicating
sponse to this pursuit response. The oscillation period is partial adaptation to the additional time delay.
about 300 ms. The middle panel of Fig. 6 is a represen-
tative eye velocity response a few trials after TAhas been
increased to 80 ms. The peak eye velocity has increased alter the visual feedback need not differentiate between
to nearly 6Oo/s and the oscillation period is approxi- these two classes of models.
mately 600 ms. After three days of three-hour training Adaptation to such a change in delay means that the
sessions with TA fixed at 80 ms, with randomized target neural delay correlate can be reconfigured. One possibil-
speed and direction, a typical eye velocity record was se- ity is the connection of additional neurons in series to pro-
lected and appears on the lower panel of Fig. 6. Evidence long compensatory delay time T,. Alternatively, in the
supporting an adaptive system is shown by a decreased spirit of our Fig. 3(e) and the partial fraction expansion
oscillation period of about 400 ms. example earlier, several parallel-path synaptic strengths
Our observations for the pursuit system's initial re- might be altered to effect a compensatory delay change in
sponse to an acute increase in time delay are similar to the controller.
those described recently by Goldreich and Lisberger [23].
They increased delay by reducing target intensity and by V . DISCUSSION A N D CONCLUSIONS

an electronic means similar to that described above. In Neural networks have appeared in conjunction with
their experiment, the oscillation period of the monkey adaptive control as early as 1960 with the paper by Wid-
pursuit response increased monotonically with added de- row and Hoff [25]. The ADELINE (ADAptive LInear
lay in the feedback loop. NEuron) pursued a least mean square algorithm based on
They claim that the parallel feedforward class of models performance feedback to form the "best" linear switch-
can reproduce this effect but that the efferent feedback ing circuit. Although this paper does not mention mech-
class cannot. We have run similar simulations on both anisms governing adaptive control, the topic has been re-
classes of models and have confirmed their observation viewed recently by Astrom [26]. Instead we focus on
for simulations where the parameters are adjusted so that application of neural nets and dynamical systems theory
the eye shows a pronounced oscillation at the nominal de- to visual tracking system neurobiology.
lay. However, when the simulation is adjusted to produce This paper has emphasized a role for internal models of
a more damped response (which fits many animals that plant dynamics and loop delay in compensation. We have
we have observed) both models make similar predictions. illustrated that these models may be either explicit
From the equivalence of the linear controllers shown in [lumped feedback, Fig. 3(d)] or implicit [parallel feed-
Fig. 3(d) and (e) it should be clear that experiments which forward, Fig. 3(e)]. The lumped-feedback controller rep-
D E N 0 cl 01.: N E U R A L NETWORK ORGANIZATION OF VISUAL PURSUIT SYSTEM 91

resentation used cancellation to explicitly recreate a cen- [9] Program CC-Computer-aided control system design package, Syst.
Technol;, Inc. Hawthorne, CA 90250.
tral signal corresponding to target position or velocity. [IO] K. J . Astrom, ”A Simnon tutorial. CODEN: LUTFDZ/(TFRT-
This strategy for the oculomotor system is attributed to 3168);’ Lund Inst. Technol., Lund, Sweden, 1982.
Young et al. [27]. We use these two extremes to motivate [ 1 I] C. Rashbass, “The relationship between saccadic and smooth track-
ing eye movements,” J . Physiol., vol. 159, pp. 326-338, 1961.
descriptions of compensation schemes which help provide 1121 J . Pola and H. J . Wyatt, “Target position and velocity: The stimulus
fast accurate tracking in the face of significant time de- for smooth pursuit eye movements,” Vision Res., vol. 20, pp. 523-
lays. 534, 1979.
131 D. A. Robinson, J. L. Gordon, and S . E. Gordon, “A model of the
Linear system theory is not wholly adequate to analyze smooth pursuit eye movement system,” B i d . Cybern., vol. 5 5 , pp.
the smooth pursuit system. It is used, in part, because it 43-57, 1986.
is a familiar and tractible analytic tool for this extension 141 S . G. Lisberger and L. E. Westbrook, “Properties of visual inputs
that initiate horizontal smooth pursuit eye movements in monkeys,”
into dynamical neural networks. Lisberger et al. [21] and J . Neurosci., vol. 5 , pp. 1662-1673, 1985.
Robinson er al. [ 131 have found nonlinear effects in mon- 151 D. C. van Essen, J . H . R. Maunsell, and J . L. Bixby, “The middle
keys at velocities as low as 5 ” / s . Our simulations and temporal visual area in the macaque: Myeloarchitecture, connections,
functional properties, and topographic organization,” J . Comp. Neu-
experimental data show that linear model parameters (cal- rol., vol. 199, pp. 293-326, 1985.
culated from our animal’s pursuit) provide adequate pre- 161 E. L. Keller, “Accomodative vergence in the alert monkey,” Vision
dictions of pursuit response if target velocities remain in R e s . , vol. 13. pp. 1565-1575. 1973.
171 D . A. Robinson, “The mechanics of human saccadic eye move-
the 20” /s and less range. Over a broader range of inputs, ment,” J . Neurophys., vol. 174, pp. 245-264, 1964.
the smooth pursuit system behaves nonlinearly and there 181 L. Young and L. Stark, “Variable feedback experiments testing a
are many ways one can seek to model these effects. The sampled data model for eye tracking movements,” IEEE Trans. Hu-
man Fucrors Electron. vol. HFE-4, pp. 28-51, 1963.
Lisberger et al. model directly determines parameters 191 D. A. Robinson, “Models of oculomotor neural organization,” in
which govern nonlinearity. However, their controller The Control of Eye Movements. P. Bach-Y-Rita and C . C. Collins,
seems less rich than those of Robinson er al. or ours since Eds., New York: Academic, 1971, pp. 519-538.
[20] D. A. Suzuki and E. L. Keller, “The role of the posterior vermis of
the linearized controller is restricted to multiple poles at monkey cerebellum in smooth pursuit eye movement control. 11. Tar-
the origin. get velocity-related Purkinjie cell activity,” J . Neurophysiol., vol.
Our pursuit system model of Fig. 3(d) permits a direct 59, pp. 19-40, 1988.
[21] S . G. Lisberger, E. J . Morris, and L. Tychsen, “Visual motion pro-
correspondence between internal models and system be- cessing and sensory-motor integration for smooth pursuit eye move-
havior. Adaptive responses by the controller to changes ments,” Ann. Rev. Neurosci., vol. 10, pp. 97-129, 1987.
in system dynamics [which follow directly from Fig. 3(d)] 1221 J . G . May, E. L. Keller, and D. A. Suzuki, “Smooth-pursuit eye
movement deficits with chemical lesions in the dorsolateral pontine
are readily reexpressed for the parallel realization of Fig. nucleus of the monkey,” J . Neurophysiol., vol. 59. pp. 952-977,
3(e). Neural firing rate data was offered which was com- 1988.
patible with a parallel feedforward realization. Evidence [23] D. Goldreich and S . G. Lisberger, “Evidence that visual inputs drive
oscillations in eye velocity during smooth pursuit eye movements in
supporting partial adaptation to increased time delay was the monkey,” Soc. Neurosci. Absrr., vol. 13, p. 170, 1987.
given. In fact, since 100-300 ms delays from sensory in- [24] F . A. Miles and S . G. Lisberger, “Plasticity in the vestibulo-ocular
put to motor response are so common one may expect de- reflex: A new hypothesis,” Ann. Rev. Neurosci., vol. 4, pp. 273-
299. 1981.
lay compensation in biologic motor control to be omni- [25] B . Widrow and M. E. Hoff, “Adaptive switching circuits,” presented
present. These may be exciting grounds for interaction at IRE WESCON Convention Record, pp. 96- 104, 1960.
between mathematical possibility, engineering plausibil- [26] K. J . Astrom, “Adaptive feedback control,” Proc. IEEE, vol. 75,
pp. 185-217, 1987.
ity, and experimental neurophysiology. [27] L. Young, J . Forster, and N. Van Houtte, “A revised stochastic sam-
ple data model for eye tracking movements,” presented at Fourth Ann.
NASA-Univ. Conf. Manual Contr., Univ. Mich., Ann Arbor, MI,
REFERENCES pp. 489-508, 1968.

[ I ] T. Kohonen and E. Oja, “Fast adaptive formation of orthogonalizing


filters and associative memory in recurrent networks of neuron-like
elements,” Biol. Cybern., vol. 21, pp. 85-95, 1976.
[2] J . A. Anderson, “Cognitive and psychological computation with
neural models,” IEEE Trans. Syst., Man, Cybern., vol. SMC-13, pp. D. Curtis Deno (M’78) was born in Schenectady,
799-815, 1983. NY, in 1954. In 1976 he received the B S degree
[3] J. J . Hopfield and D. W . Tank, “Computing with neural circuits: A in electrical engineering and biomedical engineer-
model,” Science, vol. 233, pp. 625-633, 1986. ing and the M.S. degree i n biomedical engineer-
[4] S . C. Cannon and D. A. Robinson, “An improved neural-network ing from Rensselaer Polytechnic Institute, Troy,
model for the oculomotor system: More realistic neuron behavior,” NY. He received the M.D. degree from Hahne-
Biol. Cybern., vol. 53, pp. 1-16, 1985. mann University, Philadelphia, PA in 1980.
[ 5 ] A. G . Barto, R. S . Sutton, and C. W . Anderson, “Neuronlike adap- He completed a one-year internship with the
tive elements that can solve difficult learning control problems,” IEEE Department of Medicine, Albany Medical Col-
Trans. Syst., Man, Cybern., vol. SMC-13, pp. 834-846, 1983. lege, Albany, NY in 1981 and subsequently did
[6] H. Sompolinski and I . Kanter, “Temporal association in assymetric three years as an NIH postdoctoral fellow in phys-
neural networks,” Phys. Rev. Lerr., vol. 57, pp. 2861-2864, 1986. iology, including one year of clinical trauma research, at Albany Medical
[7] E. L. Keller and N. S . Khan, “Smooth-pursuit initiation in the pres- College Since 1984 he has been working on the Ph D degree in electrical
ence of a textured background in monkey,” Vision Res., vol. 26, pp. engineering from the University of California, Berkeley, CA, where he has
943-955, 1986. advanced to candidacy. His major interests include. control systems, ma-
[8] A. Fuchs and D. Robinson, “A method for measuring horizontal and chine and biological vision, neural networks, dynamical systems theory,
vertical eye movements chronically in the monkey,” J . Appl. Phys- real-time control, neurobiology, physiology, and medicine.
id,, vol. 21, pp. 1068-1070, 1966. Dr. Den0 is a member of Sigma Xi and Alpha Omega Alpha.
1' 2

You might also like