Transcript Michael Arbib: CS564 - Brain Theory and Artificial Intelligence

Michael Arbib: CS564 - Brain Theory and Artificial Intelligence
University of Southern California, Fall 2001
Lecture 20. Saccades 1
Reading Assignments:
TMB2, Section 6.2.
Prepare for Next Time:
The NSL Book
The Modular Design of the Oculomotor System in Monkeys
Peter Dominey, Michael Arbib, and Amanda Alexander
Supplementary Reading in the NSL Book:
Crowley-Arbib Saccade Model
M. Crowley, E. Oztop, and S. Marmol
Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 20. Schemas 1
1
Neurophysiological Data on Control of Saccades
Saccades — lasting from 15 to 100 msec. — bring the eyes
rapidly to foveate on some target.
David A. Robinson 1964 showed that a saccade involved
 an initial pulse of force to drive the eye to its new position
Excitatory Burst Neuron (EBN): About 5 msec before a saccade in its
on-direction, the activity EBN(t) of an EBN jumps to a high frequency
which is maintained until about 10msec before the end of the saccade,
when it drops to the resting level appropriate to its new q.
During a saccade in the off-direction, EBN(t) is 0.

followed by a maintained force to hold the eye in its new position
When the gaze is fixed, a Tonic Neuron (TN): has firing rate
TN(t) = K[q - qo]+
where [x]+ = max(x,0).
K varies from neuron to neuron.
Motor neuron activity sums burst and toinc activity:
Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 20. Schemas 1
2
Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 20. Schemas 1
3
Excitatory burst neurons (EBNs) are thought to drive
the burst of activity in ipsilateral motoneurons.
Inhibitory burst neurons (IBNs) inhibit the contralateral motor neurons,
thus seeming to control the pause in motoneuron firing during
movements in the off direction.
Tonic neurons (TNs) exhibit regular firing at a rate that is related (in a
nonlinear way) to eye position during and after saccades.
The gain of burst cells is very large,
to keep the duration of
saccades as small as possible by generating a high velocity movement
when any appreciable motor error exists.
Omnipause neurons (OPNs) fire continuously save that they pause
during saccades. They inhibit burst cells and prevent them from
generating a saccade until released.
Another group of burst cells - trigger cells (TRIGs) - is thought to play a
role in triggering a saccade.
Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 20. Schemas 1
4
Hypotheses about saccade control
When the omnipause neurons fire (between saccades),
they inhibit the burster neurons (both EBNs and IBNs),
thus blocking the burst drive to the motoneurons.
For a saccade, the trigger cells silence the omnipause neurons; while
the excitatory burst neurons receive a signal to fire, and thus
raise the firing rate of ipsilateral motoneurons and
 via the IBNs, silence the contralateral motor neurons

for the appropriate duration for the length of the saccade.
Robinson 1975 suggested that the tonic neurons (TN) integrate the firing of
the EBNs:
TN(t) = k  t EBN(t)dt,
to obtain the desired eye position to maintain the gaze after the saccade has
occurred.
Cannon et al.offer a plausible neural-network model of the neural integrator
but the model has not yet been verified experimentally.
Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 20. Schemas 1
5
Control Systems Models of Saccade Generation
van Gisbergen, Robinson, and Gielen 1981 model the saccade generator
by a servomechanism.
Desired eye position (Ed) has subtracted an efference copy/corollary
discharge signal (E') replicating eye position (supplied by a neural
integrator NI)
to yield a motor error signal em which is input to the burst generator (B).
Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 20. Schemas 1
6
Burst neurons B send their burst directly to the motoneurons
(MN), and also indirectly through an integrator path (T, the
neural integrator NI) which has a recurrent collateral for the
"positive feedback" which provides the integration.
B is also coupled in a loop of mutual inhibition with the omnipause neurons
(P). Between saccades, firing of P inhibits all activity in B.
However, trigger input can inhibit P, thus freeing B from the tonic inhibition.
And once B is activated, it can "latch" P off for the duration of the saccade.
The saccade ends when em is reduced to 0, and inhibitory bias from the pause
neurons is restored.
Ed (and thus em) plays no further role until another saccade is triggered.
During a saccade, the MN output exhibits a pulse step in its discharge, R(t),
caused by the transient burst B(t) riding atop the eye position signal E'.
The eye is held in position by the output from the neural integrator T to the
motor neuron MN.
Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 20. Schemas 1
7
Push-pull nature of the control of burst neurons
See text for details.
Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 20. Schemas 1
8
The Many Oculomotor Systems
Each can be studied in isolation, but in normal conditions several of
these systems will work together to ensure coordinated movements:
cooperative computation at the level of motor control.
The vestibulo-ocular reflex (VOR) uses
the semicircular canals to monitor
accelerations of the head; relayed via the
neurons of the vestibular nucleus to
brainstem circuitry which causes the eyes
to compensate for head movements thus
yielding (approximate) constancy of gaze.
Such a system is of great importance in its
own right. It also plays an important role
in eye movement.
Hierarchical control: information is passed to several maps to apportion the
contributions of different movements to the overall adaptive shift of gaze as if there
were "lookup tables" for the relative contributions of eye and head.
Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 20. Schemas 1
9
The pursuit system has the task of matching the velocity of the target.
The saccade system must work with the pursuit system to track an
object adjusting position with a saccade whenever the eye moves too
far off target.
Retinal velocity even affects the saccadic system, allowing the eye to
jump into the extrapolated position of the target. Conversely, there may
be saccadic suppression to tell the pursuit system to ignore the apparent
target velocity during the saccade.
Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 20. Schemas 1
10
Dynamic Remapping in superior colliculus
Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 20. Schemas 1
11
Deeper layers of monkey superior colliculus contain cells that
only fire before saccades.
It is probably only the "center of gravity" of a peak of SC
activity which is important in coding saccadic size and direction, though burst
frequency of the ensemble of active collicular neurons affects saccade
velocity.
McIlwain 1982 found that microstimulation in the intermediate gray layer of
cat superior colliculus yields widespread synaptic activation of the layer.
Yet such stimuli evoke saccades whose metrics seem to depend primarily on the
location of the stimulating e
lectrode.
McIlwain postulates that the spatial densities of the cells projecting to vertical
and horizontal generators of the saccadic system vary systematically beneath
the retinotopic collicular map.
Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 20. Schemas 1
12
Sensor Fusion in Superior Colliculus
There are superior colliculus neurons that burst prior to saccades in
response to visual targets that also burst before saccades in response to
auditory targets.
Somatosensory input is also mapped into superior colliculus. The
somatosensory input depends on the position of body and limbs, while
position in auditory input is time-coded via the difference of arrival of
signals at the two ears.
Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 20. Schemas 1
13
From Schemas to Neural Networks - A Case Study: Modeling the
Control of Saccades
Retinotopic Mapper
Eye
Plant
Superior
Colliculus
RETINA
Brainstem Saccade Burs t Generator
Target Memory
Remapping
Retinotopic Mappe r
Eye
Plant
RETINA
Classical Model
of Reflex
Saccades
Superior
Colliculus
Brainstem Saccade Burs t Ge nerator
Expanding the
Model to Bring in
Cortical
Functions
Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 20. Schemas 1
14
Remapping of Maps in Superior Colliculus
Existing visual targets in superficial layers get remapped
to deep layers when the eye moves.
Mays and Sparks 1980, using trials in which an intervening saccade
changed the position of the eyes after a brief visual target had been
extinguished, discovered quasi-visual (QV) cells the location of whose
activity, even if the eyes had moved after the target disappeared,
represented the current retinotopic position of the remembered target.
Two processes of remapping are required
One transforms the auditory position code into a retinotopic map of the
auditory input in superficial layers of the auditory cortex.
This remapping depends on eye (and, in cat and monkey, pinna) position.
 The second translates the remembered input map by the extent of the
saccade. Dominey and Arbib (1992) adapted a model of Droulez and
Berthoz (1991) to use corollary discharge from eye movements to control
the remapping

Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 20. Schemas 1
15
q qdelay
FEF
PP
FOn
PPct r
ms
Peter Dominey & Michael Arbib:
Cerebral Cortex, 2:153-175
switch
Filling in the Schemas: Neural
Network Models Based on
Monkey Neurophysiology
qv
sm
vm
FEF
vs
PP
MD
VisCx
sm
CAUDATE
Vis Cx
SC
CD
TH
LGN
vm
SNr
SNR
vs
SG
sm
q qdelay
Develop hypotheses on Neural
Networks that yield an equivalent
functionality:
mapping schemas (functions) to the
cooperative cooperation of sets of
brain regions (structures)
FEFvs
FEFms
SC
vs
ms
qv
FOn
wta
eye movement
FEFvs
FEFms
Brainstem
Saccade
Generator
Retina
VisInput
Work at USC with Nicolas Schweighofer
Modeling the role of cerebellum in adapting the size of
saccades to compensate for
muscle linearities
 effects of prisms

Relates to experiments of Mickey Goldberg
Builds on earlier studies with Peter Dominey (discussed later in these
lectures) on
Dynamic remapping and working memory
 Role of basal ganglia in learning conditional and sequential behavior.

Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 20. Schemas 1
17
Scudder's model
for the saccade generator
addresses the issue of spatial
coding in SC: SC is intimately
related with the pulse generators:
the projection from SC to long lead bursters is
monosynaptic,
 while the latency from deep SC activity to
motoneurons is polysynaptic.

The output of SC is integrated by a neural
integrator composed of long lead burster neurons
(LLBN)
the integration is shown schematically by selfexcitation with unit weight

The retinotopic code of saccade size in SC is
converted to an intensity code by differentially
weighting the output of SC neurons according to
the eccentricity of the neuron's movement field.
Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 20. Schemas 1
18
The discharges of cells simulated by this model:
The first event is the onset of the superior colliculus
discharge
which simultaneously begins to charge the
LLBN/integrator and
to inhibit the OPN by a long latency indirect pathway.
As excitation from the LLBN/integrator discharge
increases and
the inhibition from the OPN decreases, the net EBN input goes from
inhibitory to excitatory.
As the EBN fires, the IBN fires and further inhibits the OPN, causing a
regenerative increase in EBN and IBN discharge rate.
The firing of the IBN also inhibits the LLBN/integrator, causing it to
begin to be discharged. Progressively, the
LLBN/Integrator firing
rate decreases until the inhibition of the OPN by the IBN no longer
exceeds the excitatory bias of the OPN.
The OPN begins firing, and the burst ends. The model matches the
physiological data for saccades up to 15°.
Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 20. Schemas 1
19
A problem with standard models of saccade generators:
when a movement is required in an oblique direction
the durations of the saccades generated by uncoupled horizontal and
vertical saccade generators are different, whereas the duration of both
components of an oblique saccade are equal in the animal.
Scudder showed how to couple a horizontal and vertical burst generator
pair via the omnipause neurons alone to obtain a model which yields
realistic oblique saccades.
Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 20. Schemas 1
20