Greg Detre
Thursday, 18 October, 2001
Dr J Iles, St Hughes, week 2
The parietal cortex has been
traditionally labelled an �association� area, since it seems to function
somewhere intermediate between sensory and motor functioning, involving and
integrating both types of signals. It seems to play a vital role in the visuomotor
stream (following Goodale & Haffenden�s (1998) terminology) used for
directing action, often at a subconscious level. This seems to involve the
representation and transformation between various coordinate frames of
reference, in what Andersen et al (1997) terms a �multimodal representation of
space�. More controversially, the parietal cortex also seems to contain
neuronal activity relating to attention, intention and decision.
The major areas of the parietal
cortex that are usually considered are: areas 7a and 7b, the lateral
intraparietal area (LIP), the medial superior temporal area (MST � further
subdivided into dorsal (MSTd) and lateral (MSTl) areas) and the ventral intraparietal
area (VIP). All of the areas are strongly interconnected via cortico-cortical
projections.
Broadly, LIP has been widely
implicated in saccadic movements, with strong direct projections from
extrastriate visual areas and projections to prefrontal cortex, the caudate
nucleus and the superior colliculus, all of which are areas concerned with saccadic
eye movements). MST seems highly involved in motion processing, with large RF
neurons selective for expansion/contraction, rotation and spiraling (and
sometimes more than one at a time). Area 7a has large bilateral fields, with
strong cortical connections to other visual areas, as well as the
parahippocampal gyrus and cingulate cortex. Area 7b and VIP are closely tied in
with the somatosensory system, and to a lesser extent, vision.
Skilled forelimb movements appear
to have originated early in tetrapod evolution, possibly as early as the
divergence between amphibians and amniotes (Iwaniak & Whishaw (2000), On
the origin of skilled forelimb movements).
The primary motor cortex is no
longer seen as a simple somatotopic motor representation, but rather a
multitude of representations that overlap, allowing the cortex to organise
combinations of movements for specific tasks. Movement-related neurons in the
premotor areas may fire during movements related to specific tasks and not
others to encode a more global feature, such as set-related neurons which are
active in the absence of any overt behaviour, e.g. during a delay between task
instructions and execution. The presupplementary motor area is active during
the learning of a behaviour, but becomes less active as learning progresses,
with activity in the supplementary area eventually ceasing when the behaviour
becomes automatic. Thus, the hierarchy of motor control gives rise to a hierarchy
of task features. Parts of the parietal cortex, together with the motor areas,
are heavily involved in the planning and execution of voluntary movements,
producing motor programs from the coordinate frames in which the external
environment is represented. Efference copies of motor commands, probably
generated in the frontal lobes, probably provide information to the posterior
parietal cortex about body movements, all coded in different coordinate frames,
which may be employed in subtracting eye movements from visual signals to
calculate direction of heading, for instance (Andersen et al, 1997 � see
below).
Goodale & Haffenden (1998)
argue that vision for perception and vision for action are mediated by separate
neural mechanisms. They employ a wide range of psychological and psychophysical
evidence, as well as dissocations to show that 'what we think we see is not
always what guides our actions'.
They cite the case of DF, a young
woman who developed a profound visual form agnosia following carbon
monoxide-induced anoxia. Even thought DF's low-level visual abilities remained
reasonably intact, she can no longer recognise object on the basis of form, or
the faces of freinds and relatives, nor identify even simple geometric shapes.
She is perfectly able to perform these actions using voices or touch though.
However, DF's hand and finger movements seem almost unimpaired, even when
picking up objects she cannot identify or recognise. She rotates her hand and
write quite normally, her hand opens to grasp at the right size. It appears as
though DF's visual system is no longer able to deliver perceptual information
abotu the size, shape and orientation of objects, yet the visuomotor systems
that control the programming and execution of visually-guided actions remain
sensitive to these same object features.
Moreover, there is evidence that
patients with damage to other visual areas in the cerebral cortex, e.g. the
superior regions of the PPC have the opposite behaviour - they cannot use
visual information to rotate their hand or scale finger distance when reaching,
though they are perfectly able to describe the size or orientation of objects
in that part of the visual field.
Goodale and Milner (1992) have
propposed that these two streams correlate with the 'dorsal' and 'visual'
streams identified in the cerebral cortex of the monkey. The dorsal stream
travels from the primary visual cortex to the posterior parietal cortex, and
which seems tied in with visuomotor functions.
Andersen et al (1997) argues for a
wide range of highly important functions in the parietal cortex. Foremost among
these is an abstract multi-modal distributed representation of space, combining
vision, somatosensation, audition, and vestibular sensation, which �can then be
used to construct multiple frames of reference to be used by motor structures
to code appropriate movements�, as well as selecting stimuli and helping to
plan movements. This fits in with Goodale & Haffenden�s description of the
functions of the visuomotor system, and aligns with it in neuroanatomical terms
too.
Andersen claims that areas 7a and
LIP use their eye position and retinal input signals to represent the location
of a visual target with respect to the head, a 'head-centred reference frame'.
He concedes that 'intuitively one would imagine that an area representing space
in a head-centred reference frame would have receptive fields that are anchored
in space with respect to the head', but proposes instead that instead a highly
distributed pattern is used to uniquely specify each head-centred location in
the activity across a population of cells with different eye position and
retinal position sensitivities. Indeed, he argues that 'when neural networks
are trained to transform retinal signals into head-centred coordinates by using
eye position signals, the middle-layer units that make the transformation gain
fields similar to the cells in the parietal cortex (Zipser & Andersen,
1988)'.
These �gain fields� underly
Andersen�s entire computational model of the parietal cortex. The idea is that
one component of a (local) frame of reference is stripped away or subtracted,
e.g. eye movements, to give a higher-order frame of reference (e.g. head-centred),
by shifting the receptive fields of the cells, so that the inputs the cells are
receiving are as they would have been had that dimension been kept static. By
using the gain field mechanism, a variety of modalities in different coordinate
frames can be integrated into a distributed representation of space. In this
way, information is not collapsed and lost - for instance, if the gain field
mechanism is used to produce a head-centred frame from retinal position and eye
position, the eye-centred coordinates could be read out by another structured -
the two components have not been converged - it's almost like shifting all the
information in a spreadsheet one column along. Lesions to the posterior
parietal cortex give rise to spatial deficits in multiple coordinate frames.
This could be because many coordinate frames might conceivably representable in
the same population of neurons. Or it could simply be that the different
coordinate frames exist in close proximity to one another and so would all be
affected at the same time.
The idea of gain fields is best
understood with reference to their explanation of how MST might be able to
calculate direction of heading from visual signals, or �'computing the
direction of self-motion in the world based on the changing retinal image'.
Finding one�s heading based on visual information is relatively easy if one is
facing in the right direction, since it simply involves using the centre of the
expanding visual motion generated by self-motion as the direction of heading
(Gibson, 1950). In order to recover the direction of heading even when we are
fixating/tracking an object that is not directly ahead of us though, we have to
decompose the resulting optic flow field into a) the movement of the observer
(expanding field) and b) eye rotation (linearly moving field). Andersen relies
on Royden et al�s (1992) suggestion that an efference copy of the pursuit
command may be crucial here. Handily, MSTd contains cells selective for one or
more of the following: expansion-contraction, rotation and linear motion
(Saito, 1986). However, it appears that MSTd is not decomposing the optic flow
into channels of expansion, rotation and linear motion - Andersen produced a
spiral space with expansion on one axis and rotation on another, and found that
disappointingly few of the MSTd neurons had tuning curves aligned directly
along these axes. Interestingly though, the MSTd neurons displayed a high
degree of position and size invariance, as well as form/cue invariance. The
MSTd cells seems to convey 'the abstract quality of a pattern of motion, e.g.
rotation', which may be important in analysing optic flow by gathering
information from any part of the visual field. MST may use cells sensitive to
motion pattern in combination with the pursuit eye movement signal it receives
to code direction of heading. They found that many MSTd neurons shift their
receptive fields during eye pursuit movements to more faithfully code the
direction of heading than the focus of expansion on the retina. For instance,
when viewing an expanding pattern while making a pursuit movement towards, say,
the left, the retinal position of the focus shifts left, which many
expansion-selective MSTd neurons compensate for by shifting their receptive
fields to the left (and often vice versa for rightward movements). In
Andersen�s words:
When the eyes move, the focus tuning curve of
these cells shifts in order to compensate for the retinal focus shift due to
the eye movement. In this way MSTd could map out the relationship between the
expansion focus and heading with relatively few neurons, each adjusting its focus
preference according to the velocity of the eye.
This pursuit compensation is
achieved by a non-uniform gain and distortion applied to different locations in
the receptive field. Andersen acknowledges that Perrone & Stone's (1994)
and Warren's (1995) models are similar, but require more neurons for separate
heading maps for different combinations of eye direction and speed (rather than
just eye movement). Andersen goes so far as to say that MSTd may compensate
spatially for the consequences of eye movements for all patterns of motion.
Andersen applies the idea of gain
fields to show how eye-centred, head-centred, body-centred and world-centred
representations can be built up, conceivably even on top of one another, in the
parietal cortex. This requires integrating retinal signals, eye movements,
proprioceptive input (especially from the neck), auditory information
(intra-aural time, intra-aural intensity and spectral cues from both ears),
body motor outputs and vestibular input.
Furthermore, he claims that neural
network models can illustrate methods employing gain fields to transform
between coordinate frames. For example, Zipser & Andersen (1988) showed
that when 'retinal position signals are converted to a map of the visual field
in head-centred coordinates, the hidden units that perform this transformation
develop gain fields very similar to those demonstrated in the posterior
parietal cortex', and that the activities found for posterior parietal neurons
could be the basis of a distributed representation of head-centred space. In
Xing et al's (1995) model, which takes in head-centred auditory signals and eye
position and retinal position signals as input, and whose output codes the
metrics of a planned movement in motor coordinates, the middle layers develop
overlapping receptive fields for auditory and visual stimuli and eye position
gain fields. It is interesting that the visual signals also develop gain
fields, since both the retinally based stimuli and the motor error signals are
always aligned when training the network and, in principle, do not need to use
eye position information. However, the auditory and visual signals share the
same circuitry and distributed representation, which results in gain fields for
the visual signals.
More controversially, Andersen
claims that the PPC also contains circuitries that appear to be important for
shifting attention, stimulus selection and movement planning. Patients with
lesions to the PPC have difficulty shifting their focus of attention (Posner et
al, 1984). It now seems that visual responsiveness of parietal neurons is
actually reduced at the focus of attention (Robinson et al, 1995), while
locations away from the focus of attention are more responsive, apparently
signaling novel events for the shifting of attention.
Gnadt & Andersen (1988) have
shown that activity in cells primarily in LIP (coding in oculomotor
coordinates) precedes saccades. This activity is also memory-related, e.g.
lighting up when a monkey is remembering the location of a briefly-flashed
stimulus and, after a delay, made a saccade to the remembered location.
Glimcher & Platt required an animal to attend to a distractor target, which
was extinguished as a cue to saccade to the selected target, thus separating
the focus of attention from the selected movement. For many of the cells, the
activity reflected the movement plan and not the attended location, although
the activity of some cells was influenced by the attended location. Andersen
thinks that these and other studies suggest that a component of LIP activity is
related to movements that the animal intends to make.
Mazzoni et al (1996) used a
delayed double-saccade experiment to try and distinguish whether the memory
activity was primarily related to intentions to make eye movements or to a
sensory memory of the location of the target. They found both types of cells,
with the majority of overall activity being related to the next intended
saccade and not to the remembered stimulus location. This did not necessarily
lead to execution of the movement, since the animals could be asked to change
their planned eye movements during the delay period in a memory saccade task,
and the intended movement activity in LIP would change correspondingly
(Bracewell et al, 1996). If it could be shown that the activity is related to
the type of movement being planned, it would be a strong indication that the
activity is intention-related. Bushnell et al (1981) recorded from PPC neurons
while the animal programmed an eye or reaching movement to a retinotopically
identical stimulus. They claimed that the activity of the cells did not
differentiate between these two types of movements, indicating that the PPC is
concerned with sensory location and attention and not with planning movements.
However, when Andersen et al repeated the experiment, they found that 2/3 of
cells in the PPC were selective during the memory period for whether the target
requires an arm or eye movement. Andersen considers Duhamel et al, 1992
(similar to Gnadt & Andersen, 1988) and Kalaska & Crammond, 1995 as
studies in which their theory that the memory-related activity in the PPC
signals the animal's plan to make a movement could explain the results. Thus,
when stimulus-related activity comes into the parietal cortex, it can sometimes
invoke more than one potential plan, e.g. both eye and limb movements, even if
the limb movement is not executed.
It seems clear that the parietal
cortex plays a variety of roles, relating sensory and motor functions. The
majority of the debate centres around:
1.��� The nature and
variety of the representations and transformations of coordinate reference
frames encoded in the parietal cortex. Graziano and Gross (1998) argue that
there is no single spatial coordinate system, but that the PPC carries the �raw
data� necessary for other brain areas to construct spatial coordinate systems.
The highest levels of spatial processing are deep within the motor system. They
seem to suggest that the frontal lobes have more of a spatial role to play than
PPC.
2.��� Whether cells in
the parietal cortex signal intention and the type of movement to be performed
(as opposed to being a sensory memory of the location of a target perhaps, or
less likely, precursors for actual execution of movement), as Andersen claims.
In contrast, Bushnell et al (1981) have claimed that the activity PPC cells
while the animal programmed an eye or reaching movement to a retinotopically
identical stimulus did not differentiate between these two types of movements,
indicating that the PPC is concerned with sensory location and attention and
not with planning movements.
However, all of the evidence
supports a claim for a dichotomy of visual processing. Burr et al divide them
into one for conscious perception (more plastic and subject to spatial
distortion) and the other for the control of action. This is similar to Goodale
& Milner�s (1992) distinction, but it is probably more helpful to think in
terms of visuoperceptual and visuomotor functions, rather than phenomenology.