Download similar cortical mechanisms for perceptual and motor learning

Review
TRENDS in Neurosciences Vol.27 No.8 August 2004
Viewing and doing: similar cortical
mechanisms for perceptual and motor
learning
Rony Paz1,2, Steven P. Wise3 and Eilon Vaadia1,2
1
Department of Physiology, Hadassah Medical School, The Hebrew University, Jerusalem 91120, Israel
The Interdisciplinary Center for Neural Computation, The Hebrew University, Jerusalem 91904, Israel
3
Laboratory of Systems Neuroscience, National Institute of Mental Health, Bethesda, MD 20892-4401, USA
2
Historically, different groups of researchers have investigated the mechanisms of perceptual learning and
motor learning. For sensory cortex, neurophysiological
and psychophysical findings have linked changes in
perception with altered neuronal tuning properties.
However, less information has been forthcoming from
motor cortex. This review compares recent findings
on perceptual and motor learning, and suggests that
similar mechanisms govern both. These mechanisms
involve changes in both the center of neuronal tuning
functions and their width or slope. The former reflects
the values of the sensory or motor parameters that a
neuron encodes, and the latter adjusts the encoding
sensitivity. These similarities suggest that specific unifying principles for neural coding and computation exist
across sensory and motor domains.
Pity the plight of baseball batters: sometimes a pitch flies
straight and fast, but other times it approaches on a slow
curve. Batters must develop the motor skills needed to hit
both pitches and choose between them as the ball travels
toward them at 35 – 45 m s21. Accordingly, they must learn
to recognize which pitch is which, pronto. Time does not
permit the batter to weigh the pros and cons, the evidence
for one pitch or the other, or the need for one motor skill or
the other. The relevant knowledge is all implicit; the
memories procedural. Rapidly rotating balls curve in
flight, and so – to identify the pitch – batters detect the
pattern made by the red stitches in the ball as it spins.
That skill requires perceptual learning; the crucial batting
skills require motor learning. Yet the brain of a batter
must acquire both kinds of skill, and, if it cannot do so, that
brain needs to find a different occupation.
In trying to understand both perceptual and motor
learning, neuroscientists have adopted a ‘divide-andconquer’ strategy, in which different researchers approach
various parts of a problem. The ‘conquer’ part requires a
synthesis of findings. David Marr’s distinctions of levels
of understanding brain function [1] might help us do that.
He distinguished the level of a ‘computational theory’,
which clarifies the problem to be solved, from the levels of
Corresponding author: Rony Paz ([email protected]).
Available online 11 May 2004
‘algorithm’ and ‘implementation’, which involve solving
the problem. Because sensory and motor systems use
similar ‘hardware’ – neurons and synapses – they must
have close analogies at the level of implementation.
However, the issue is whether those similarities extend
to the algorithmic level. Such similarities would simplify
the interaction between sensory and motor systems, but
have neuroscientists obtained any empirical evidence to
support this idea?
At one level, sensory and motor systems resemble each
other closely: almost all brain areas have neurons with
‘activity fields’, also known as receptive fields, motor fields
or tuning curves (Figure 1). A population of neurons with
different activity fields can provide the basis for representing sensory or motor parameters, and as such they can
serve as the ‘primitives’ – the fundamental components –
of a neural representation. The nature of these primitives
and their locations in the brain can sometimes be inferred
by measuring how learned skills transfer to situations
other than those experienced previously. Take a simple
perceptual skill, for example, the ability to discern the
difference between bars of light oriented at slightly
different angles. Imagine that learning this skill for one
set of orientations at one place in the visual field does not
generalize (or transfer) to other orientations or to other
places in the visual field. In that case, such learning probably
depends on narrowly tuned primitives at the level of the
primary visual cortex (V1), where any given neuron
responds to a small range of locations and orientations [2].
Ahissar and Hochstein [3] extended this idea and suggested
that difficult visual skills depend on learning mediated by
‘lower’ visual areas, where narrow tuning leads to poor
generalization, whereas easy skills depend on ‘higher’ visual
areas, which have the opposite properties.
The same principles apply to the motor system. The
concept of neuronal primitives suggests that specific
computational principles exist for sensorimotor transformations [4,5] and imposes constraints on the learning of
those transforms [6,7]. As with perceptual learning,
examination of how motor skills transfer reveals something about the primitives underlying motor learning. For
example, the extensive generalization of learning seen
for smooth-pursuit eye movements of different speeds
suggests broadly tuned primitives [8], and studies of
www.sciencedirect.com 0166-2236/$ - see front matter. Published by Elsevier Ltd. doi:10.1016/j.tins.2004.04.013
Review
TRENDS in Neurosciences Vol.27 No.8 August 2004
(a)
M1
V1
A1
Preferred orientation (PO)
Preferred direction (PD)
Best frequency (BF)
Activity (spikes s–1)
(b)
Before
After
Orientation of a light bar
Acoustic frequency
Force
Movement direction
TRENDS in Neurosciences
Figure 1. Neural primitives of representation take the form of activity fields in both
sensory and motor areas. (a) The general location of the primary visual cortex
(V1, green), the primary auditory cortex (A1, magenta) and the primary motor
cortex (M1, blue). Adapted from Ref. [65]. Cells in these areas usually show
sensitivity to a low-level feature of the stimulus or movement, and usually ‘prefer’
one value of this feature over others. Such values have different names for
different systems, such as preferred orientations (PO) in V1, best frequencies (BF)
in A1 and preferred direction (PD) in M1. (b) Tuning curve of a generic cell (gray)
and two possible learning-related changes (red). The gray curve shows an the
activity of an ideal cell as a function of the measured variable, be it orientation of a
light bar (green), the frequency of a tone (magenta) or the direction of movement
(blue), with the color of the x-axis corresponding to a cortical area in (a). The red
curves indicate two types of change in the tuning of the cell that could result from
learning. The left curve indicates a change in the width of the tuning curve without
any change in amplitude of the peak or tuning value, and the right curve indicates
a shift in the tuning value without any change in either the width or peak amplitude
of the curve. The pictures on the right of each x-axis depict an improvement
in either a perceptual or a motor skill. From top to bottom: better discrimination
of line orientation, improved detection of an acoustic signals, and adaptation
of reaching movements to imposed forces to restore accurate and straight
trajectories towards a goal.
force-field adaptation suggest a reliance on primitives
that resemble the tuning curves of cerebellar Purkinje
cells [9]. Specifically, some cerebellar tuning curves for
hand velocity have two peaks: one in each of two
diametrically opposed movement directions [10]. This
property helps to account for the transfer of force
adaptation from a practiced movement direction to movements in the opposite direction [9].
This review highlights some recent findings on the
mechanisms of perceptual and motor learning, which point
to unexpected similarities at Marr’s algorithmic level.
Furthermore, whereas studies of learning-related changes
typically focus on high-order cortical areas – the so-called
association cortex – examples from primary cortical areas
are emphasized here: one motor area [the primary motor
cortex (M1)] and two sensory areas [V1 and the primary
auditory cortex (A1)].
www.sciencedirect.com
497
Learning-related changes in tuning values
The description of neuronal activity is commonly reduced
from an equation specifying the full tuning curve (Figure 1)
to a single ‘tuning value’, which commonly corresponds
to the greatest discharge rate and is referred to as the
‘preferred’ or ‘best’ value. For example, A1 neurons have a
best frequency (BF) for responding to tones, V1 cells have a
preferred orientation (PO) for responding to lines and bars,
and M1 cells have a preferred direction (PD) for reaching
movements. This compact description of neuronal tuning
is useful in many coding schemes, for example, the
population vector, in which the PD of M1 cells and their
instantaneous activity estimate movement direction.
Learning sometimes induces shifts in the ‘tuning value’,
which can be described as either attractive (a change
towards some parameter) or deflective (a change away
from some parameter). As an example of deflective
changes, after adapting to a grating of one orientation,
neurons in V1 of cats and monkeys shift their PO away
from that orientation [11]. As an example of attractive
changes, after applying intracortical microstimulation
(ICMS) to a subpopulation of neurons with tuning to one
orientation, adjacent neurons shift their PO towards that
of the stimulated ones [12]. Another attractive tuning
change occurs in V1 after pairing an orientated line
stimulus with ICMS [13]. In the auditory system, attractive BF shifts follow auditory fear conditioning at a given
frequency, acoustic discrimination learning, frequency
adaptation and application of ICMS to a subpopulation
of neurons with tuning to one frequency [14,15]. Deflective
shifts occur only rarely [15].
The results from sensory systems have some recently
discovered correlates in the motor system. Skill acquisition affects the PDs of single cells in M1 [16– 19], as does
adaptation to directional errors induced by force fields
[20,21]. In the latter experiments, monkeys practiced
reaching movements while holding a robotic arm that
imposed a complex pattern of forces. For example, one
pattern of force was proportional to the velocity of the hand
in both dimensions of a two-dimensional workspace. When
monkeys learned to reach in this new environment, M1
cells shifted their PDs in about the same way as the
muscles did. Moreover, a population of cells maintained
their new PDs (and altered activity levels) after the
monkeys readapted to moving without the imposed forces.
Interestingly, adapting to local visuomotor transformations [22] and to viscous loads [23] did not induce
consistent shifts in the PDs of M1 cells, so changes in
the tuning value have not been universally observed in
studies of motor learning.
Further, albeit indirect, evidence for changes in the
tuning value in M1 comes from the work of Classen et al.
[24]. They had subjects move their thumb repeatedly in
one direction and, after this training, applied transcranial
magnetic stimulation (TMS) to M1. In this condition, the
elicited movements systematically shifted towards the
training direction – an attractive plasticity. This result
could have come from a shift in the PDs of M1 cells,
although this hypothesis remains to be tested. Nevertheless, the occurrence of PD shifts in these instances
Review
TRENDS in Neurosciences Vol.27 No.8 August 2004
suggests fundamental similarities in sensory and motor
plasticity at the cortical level.
Which neurons change?
Another aspect of learning involves the selection of a
subpopulation of neurons involved in a change in tuning
properties. In V1, observed shifts in PO after adaptation
to one orientation occur only in cells with nearby POs.
Moreover, the larger the difference between the PO of a
cell and the training orientation, the smaller the shift
(Figure 2a) [25]. In the auditory system, the situation is
similar: maximal shifts in BF were observed for cells with
BFs close to the training BF (but not for those ‘too’ close) [26].
In the motor system, Paz et al. [22] recorded the activity
in M1 before, during and after monkeys adapted to visuomotor transformations commonly called ‘rotations’. In the
baseline condition, the monkey moved a cursor on a video
monitor in a straightforward manner, similar to the
mouse –cursor relationship on a computer. For example,
when the monkey moved its hand forward, the cursor
moved upwards (908 when 08 is to the right). During the
learning condition, the movement of the cursor was
‘rotated’ so that a hand movement in some other direction
moved the cursor upwards. For example, a 2 458 rotation
required the monkey to learn that a hand movement
forward and to the left led to upward cursor movement,
and movements of the hand in any other direction also
caused the cursor to move at a rotational transformation of
2 458. In the experiment, the monkeys learned a different
‘rotation’ each day, but saw only one target during the
training period. Before and after training, the monkey
moved the cursor to targets in eight directions from a
central location in the baseline condition. In agreement
with studies in humans involving similar transforms [27],
the monkeys showed limited generalization of learning for
Shape of the tuning curve and implications for improved
coding
Learning in sensory systems might affect the shape of
tuning curves. For example, the slope of the curve might
change at a particular point along the curve, even without
a significant change in the tuning value (the center or peak
of the curve) or the amplitude of the peak. This possibility,
which is illustrated in Figure 1b, has several implications
for neural coding. In one view, neuronal tuning curves
encode the value of a stimulus by signifying their preferred
value and the population serves as a set of basis functions.
According to this idea, decoding can occur through function approximation [28,29] – that is, approximating stimulus value from the activity of a population of neurons with
diverse tuning values and tuning-curve widths [30].
Viewed from a somewhat different perspective, tuning
(b)
(c)
30
0.4
P < 0.05
P > 0.05
15
Error
Maximal orientation shift (deg)
(a)
movement directions other than the training direction.
There was some generalization for movements within ^458
of the training direction, but virtually none for movements
in more distant directions (Figure 2b). That is, the motor
learning was largely local. Paralleling these behavioral
findings, changes in activity were observed only in neurons
with PDs close to the training direction (Figure 2c).
This specificity in the way that the representational
primitives of M1 change during motor learning could
explain the limited spatial generalization observed. Only a
selective subpopulation of cells, those with tuning values
in or near the training direction, participated in the
learning process. Movement in many different directions,
however, requires participation of neurons with a wide
range of tuning values. But neurons tuned to directions far
from the training direction did not change. Hence, as in the
sensory system [2], learning did not automatically transfer
from a training place or direction to distant ones.
0
30
60
90
Post-learning trials
1
2
3
4
5
6
0.2
0.0
–0.1
315
–15
Orientation difference (deg)
0
45
90 135 180 225 270
Direction (deg)
Difference in normalized rates
498
0.4
0.2
0.0
–180:–150:–90: –30: 30:
–150 –90 –30 30 90
90: 150:
150 180
PD distance from learned direction (deg)
TRENDS in Neurosciences
Figure 2. Learning-related changes occur in a selected subpopulation of cells. (a) In the task used to generate these data, brief presentations of a grating stimulus (alternating dark and bright bars) caused adaptation to its orientation. For cells in the primary visual cortex (V1), shifts in their preferred orientation (PO) depended on the angular
distance of their original PO from the presented orientation (black represents significant shifts). The closer their original PO to the one presented, the more their PO shifted.
Negative values indicate attractive shifts (i.e. shifts towards the experienced orientation). This finding indicates that changes in activity occurred most prominently in a
selected subpopulation of neurons. Adapted, with permission, from Ref. [25]. (b) Monkeys adapted on a daily basis to rotational visuomotor transformations that required
them to move their hand at an angle relative to the direction of a cursor. During training, only one target appeared (indicated as 908). The plot shows the after-effects of
training for movements in eight directions, all tested after adaptation but with no ‘rotation’ (the baseline condition). After-effects were most pronounced for the training
direction (908) and decreased as a function of angular distance from it (and as a function of number of trials, shown by the different colors), indicating poor generalization
and showing that adaptation is local with respect to movement direction. Adapted, with permission, from Ref. [22]. (c) For primary motor cortex (M1), neuronal activity was
recorded before, during and after the learning shown in (b). Each value on the x-axis represents a range of angular deviations from the preferred direction (PD) of the cells.
The notations describe a range of PD differences between the value at the top and the value at the bottom, separated by a colon. Cells increased their activity in a delay
period preceding movement, but only for movements in or near the training direction and only for cells with PDs near that direction. This shows that changes occur in a
selected subpopulation of neurons and could explain the poor generalization observed in (b). Adapted, with permission, from Ref. [22].
www.sciencedirect.com
Review
499
TRENDS in Neurosciences Vol.27 No.8 August 2004
curves represent a probability distribution through which
neurons encode stimulus intensity with a degree of uncertainty included [29]. Decoding could occur through a
Bayesian-inference process, which estimates the probability of a specific stimulus based on neuronal firing rates.
Two factors have special importance in this process: the
signal-to-noise ratio and the slope of the tuning curve.
Lower noise allows a closer estimate of the veridical
stimulus value by each single observation and thus leads
to more reliable read-outs. Steeper tuning curves provide
more distinguishable firing rates at adjacent points along
the curve, increasing the sensitivity of a neuron to
stimulus values encoded on its flanks (Figure 1b).
Supporting these ideas, several studies have reported
increased neuronal sensitivity [25,31] and decreased variability [31] in V1 neurons after brief adaptation to gratings
of one orientation (Figure 3a). A1 neurons also exhibit
sharpening of their tuning curve after focal electrical
stimulation [32] and following training on a frequency
(a)
discrimination task [33]. Recently, Schoups et al. [34] have
studied changes in the slope of V1 tuning curves after
monkeys mastered a fine-grained, orientation identification task. They found that the flanks of tuning curve for
the orientation of a bar-like stimulus have a higher slope
than before the perceptual learning (Figure 3b). Thus, in
sensory systems, learning-related changes in the width of
a tuning curve could be secondary to an increase in the
slope of its flanks.
Until lately, evidence of this kind was unavailable for
the motor system. Recent studies, however, show that
cells change the strength of their directional preference
during the course of learning [16,18,19] and that this
plasticity can aid in predicting planned and performed
movements [18,19,35,36] (Figure 3c). As in sensory systems, probability-based analytical approaches show promise [37]. For example, under certain conditions, better
decoding of movement direction can be achieved by using a
Bayesian estimator than by using a population vector [37].
(b)
Slope at TO
(% change deg–1)
Mean spike count
1.6
Control
Adaptation
1.2
0.8
0.4
0.0
3.0
2.0
1.0
0
0
45
90
135
Orientation (deg)
–32
0
16
32
47
–47
–16
Preferred orientation (PO) – Trained orientation (TO)
180
(c)
(d)
∆ Mean R2
0.1
0.0
0.5
2
0.0
–0.5
–0.1
–1.0
0
10
20
30
Day of training
40
r = 0.492
P = 0.002
1
0
Change in slope
Information increase
after learning
(bits)
1.0
–90
0 90
Distance from
training direction
–5
0
5
Change in slope of tuning curve
(spike per rad/4)
TRENDS in Neurosciences
Figure 3. Learning-related changes in the shape of tuning curves. (a) After a brief period of adaptation to a grating orthogonal to the preferred orientation (PO) of the cell,
the tuning curve of this primary visual cortex (V1) cell sharpened (gray) relative to its tuning before that training (black). Sharpening of the tuning curve could increase the
sensitivity of the neural code. As the cell responded to a narrower spectrum of stimulus values, its response encoded more specifically the actual stimulus value. Adapted,
with permission, from Ref. [25]. (b) In a different study of V1 cells, monkeys were trained to identify the orientation of a small grating and showed improvement that was
specific for both stimulus location and orientation. The graph shows the slope of the flanks of the tuning curves for a population of V1 neurons, as a function of how much
the PO of each cell differed from the training orientation (TO). With training (red line), that slope increased relative to the absence of training (broken blue line), but only for
neurons with a PO within a 12–208 range from the training orientation. The increase in slope provides a larger difference in response to adjacent values of the stimulus and,
therefore, improves sensitivity and discriminability of the stimuli. Adapted, with permission, from Ref. [34] q (2001) Nature Publishing Group (http://www.nature.com/).
(c) In a study of primary motor cortex (M1) neurons, monkeys moved a cursor through neuronal activity, alone (brain control), or with their hand. R 2 values represent the
degree correlation of each cell with movement direction. The plot shows the difference D in R 2 (brain control minus hand control) as monkeys learned the brain-control
skill. The degree of directional tuning of M1 cells showed daily improvement as a function of training in the brain-control task. Adapted, with permission, from Ref. [18]
q (2002) American Association for the Advancement of Science (http://www.sciencemag.org). (d) A different study of M1 neurons [22] showed a correlation between the
information content of the cells about movement direction and the slope of their tuning curves. Information content was measured as the mutual information between
firing rate and direction of movement for single cells after learning rotational transformations. Some cells significantly increased their information content after learning
(red asterisks) but others did not (gray dots). Information content correlated with the slope of the tuning curve (main plot), but did so only for the training direction (inset).
The increased slope observed in M1 is reminiscent of the change observed in V1 in (a) and could enable finer distinctions among nearby movement directions. Adapted,
with permission, from Ref. [38].
www.sciencedirect.com
500
Review
TRENDS in Neurosciences Vol.27 No.8 August 2004
Evidence that the encoding process can also support
probabilistic approaches was recently revealed by a close
examination of neuronal activity after learning visuomotor ‘rotations’ of the sort already described. Paz and
Vaadia [38] demonstrated that M1 neurons contained
more information about direction of movement after
learning than before, and that this increase in information
content correlated with an increase in the slope of the
tuning curve (Figure 3d). Furthermore, this slope increase
was specific to the part of the tuning curve near the
training direction (Figure 3d, inset).
Why would M1 transmit more information about the
executed movement and allow improved decoding? According to the traditional view, better information promotes
efficiencies in spinal computations, leading to enhanced
performance in terms of success in achieving goals within
task constraints. In addition, however, the motor system
also transmits an internal signal, termed efference copy
or corollary discharge [39], which informs other computational networks about the generated movement. This
signal can help plan subsequent action. More importantly,
it can generate a prediction about the potential outcome
of a movement in time to provide corrections before a
movement has ended, a computation called a forward
model [40,41]. Accordingly, improved encoding of information provides as much benefit for the motor system as
for sensory systems, perhaps because, as for sensory areas,
other brain areas need to use the signals generated by
motor areas.
Perceptual learning is thought to result from an
expanded representation of the trained stimulus dimensions [42], and similar concepts have been suggested for
motor learning [24]. Results such as those discussed in this
section suggest that instead of thinking about expanded
representations in terms of the number of cells encoding a
given stimulus dimension or movement parameter, a more
subtle change in the way that cells encode stimuli, which is
reflected in tuning-curve adjustments, could account for
both perceptual and motor learning.
Contextual specificity and complexity of responses
Although many neurons in sensory areas are tuned to the
types of low-level stimulus features discussed so far, they
can also be tuned to complex interactions of low-level
features that cannot be predicted from their linear combinations, including specific visual objects [43] and natural,
complex acoustic signals [44,45]. In the motor system, an
example of this phenomenon comes from studies of
bimanual arm movements. The responses of M1 neurons
cannot be explained as a linear combination of their
responses during performance of the separate unimanual
movements that compose them [46,47]. Interestingly,
similar observations have been made for binaural interactions: many A1 neurons exhibit complex interactions,
responding maximally to specific combinations of sound
levels in the two ears [48].
Learning can also be specific to the context of skill
acquisition. This principle has long been established for
the sensory system [2]. For example, Crist et al. observed
limited transfer of perceptual learning between a spatial
bisection task and a Vernier task (Figure 4a, left), despite
www.sciencedirect.com
the fact that both require enhanced acuity for the spatial
separation between lines [49]. Similarly, V1 cells exhibit
contextual modulation in the bisection task [50]. A line
outside their receptive field affected their response to a
parallel line inside it, but only when the animal performed
the bisection task (Figure 4a, lower right), not when it
performed a fixation task (Figure 4a, upper right).
Another example of context dependency comes from the
auditory system. Ulanovsky et al. [51] recorded the activity
of single cells in A1 while cats heard sequences of two
frequencies (close to the BF of the cell). When the first
frequency was rare, cells responded more strongly to it
than to the second frequency. However, when the first
frequency was common, cells responded less strongly to it
that to the other frequency (Figure 4b). Thus, neurons in
A1 adapted in a specific contextual way; they modulated
their activity in response to the statistics of stimulus
occurrence (rare versus common), rather than exclusively
to tone frequency.
Similarly, learning new motor tasks can be context
dependent [52]. For example, evidence for dependency of
motor learning on visual context was provided by Cohn
et al. [53], who trained people to adapt to Coriolis forces in
an environment in which their whole bodies rotated. When
the participants experienced a visual-rotation illusion,
which mimicked the context of Coriolis forces, they compensated for reaching movements as if they were under the
influence of that force.
Can people learn two motor skills in parallel, but in
different contexts? If so, then it seems likely that the motor
system uses multiple controllers that can be selected and
switched according to task context [54,55]. In some situations, learning multiple skills in parallel seems to be
difficult. For example, sometimes when people try to learn
a second skill before they have fully consolidated the first,
practice on the second skill disrupts that consolidation. In
these experiments, people learn to overcome errors in
movements that are induced experimentally. The difficulty
in learning two skills in quick succession occurs both when
the induced errors depend on limb dynamics (forces) [56]
and when they depend on kinematics (position and its
derivatives) [57]. However, when errors in one direction
(e.g. clockwise errors) depend on dynamics, whereas those
in the other direction (e.g. counter-clockwise errors)
depend on kinematics, parallel learning can occur [57].
This finding suggests that, at least in the some conditions
[58], people learn dynamics and kinematics in different
contexts and depend on at least partially different neural
systems [59]. Furthermore, even within the domains of
kinematics and dynamics, parallel context-dependent
learning is possible. For kinematics, consolidation is not
disrupted when people experience opposing rotations
while performing two different tasks (e.g. a continuous
drawing task versus a reaching task [60]). For dynamics,
when two opposing force fields are presented at different
hand positions, people can simultaneously learn both
skills and switch between them [61]. Adapting to two force
fields that are based on contextual cues takes extensive
practice, and might be impossible under some circumstances [62]. In other situations, however, such as when
random and frequent switching of context occurs, people
Review
(a)
(b)
Normalized
response
8
Fixation
trials
(c)
1
f1 f1 f1f2 f1 f1 f1
2
f2 f2 f1 f2 f2 f2 f2
3
f1 f2 f1 f1 f2 f2 f1
f1 occurs at
10%
50%
90%
1
–1
0
1
Spikes s–1
Normalized
response
8
Bisection
trials
501
TRENDS in Neurosciences Vol.27 No.8 August 2004
1
–1
0
f2 occurs at
10%
50%
90%
1
Distance (deg)
Vernier
task
0
230
Time (ms)
TRENDS in Neurosciences
Figure 4. Contextual modulation of activity in primary cortical areas. (a) In a study of primary visual cortex (V1) cells, response properties were measured when monkeys
performed either of two tasks: detection of a small change in the brightness of the fixation spot (fixation trials) and a three-line bisection task in which subjects determined
the position of a line relative to parallel lines (bisection trials). (Also shown is a depiction of the Vernier task, in which a line is judged in relation to a colinear line rather
than in relation to two parallel lines, as in the bisection task.) The top right-hand plot shows the tuning curve of one cell as a function of the placement of parallel bars while
the monkey was performing the fixation task. The plot in the lower right-hand corner shows the tuning curve of the same cell for the same set of parallel bars while the
monkey performed the bisection task. The very different shape of the tuning curve shows that some V1 cells have a task-dependant response that cannot be explained
entirely by the low-level physical features of the stimulus. Adapted, with permission, from Ref. [50] q (2001) Nature Publishing Group (http://www.nature.com/). (b) In a
study of primary auditory cortex (A1) neurons, each stimulus set (upper panel) consisted of three blocks. In block 1, the lower-frequency tone (f1) was common (90%) and
the higher frequency tone (f2) was rare (10%). In block 2, these percentages were reversed, and in block 3 the two stimuli occurred equally often. The average response of a
cell to f1 appears in the middle panel, that to f2 in the lower panel. The cells responded to both stimuli more when they were rare (red) than when they were common
(blue), and therefore modulate their activity in response to the statistics of stimulus occurrence. Adapted, with permission, from Ref. [51]. (c) For the study of primary
motor cortex (M1) cells depicted in Figure 3(c), lines connecting hand-controlled preferred directions (PDs) with brain-controlled PDs (circle ends) projected onto a unit
sphere. Most cells had different PDs in the two task contexts. Adapted, with permission, from Ref. [18] q (2002) American Association for the Advancement of Science
(http://www.sciencemag.org).
can learn to reach accurately in two opposing, dynamicsdependent force fields, instructed only by nonspatial
auditory and color cues [63].
Context effects can also be observed in the activity of
M1 cells [64]. For example, Taylor et al. [18] trained
monkeys to control the movement of a cursor in a threedimensional virtual-reality environment, either by the
movement of their hand or via a brain– machine interface
that translated M1 activity into cursor movement. They
found that many cells acquired different PDs for the
two similar, yet contextually different, tasks, and that
this difference evolved systematically over many days
(Figure 4c). In a similar task, monkeys controlled a brain–
machine interface and performed reach and grasp movements [19]. Changes in the strength of directional tuning,
distribution of preferred directions and strength of
correlations were evident upon switching between braincontrolled and hand-controlled tasks. Thus, for both
perceptual and motor learning, neuronal activity reflects
not only the physical properties of the stimulus or movement, but also information about the relevant context.
Concluding remarks
William of Ockham (c. 1285– 1349), the medieval philosopher famous for formulating the principle of parsimony
www.sciencedirect.com
(‘Ockham’s razor’), would have wanted a unified account of
perceptual and motor learning. The brain must solve
problems within the constraints of its hardware: neurons
and plastic synapses. It should not be surprising, therefore, that the solutions it has found for perceptual and
motor learning resemble each other at Marr’s level of
implementation. The similarities discussed here for the
algorithmic level are more intriguing, and they also reflect
a basic fact about the brain: like the baseball batter
described in the opening paragraph, we have but one
brain, which must both do and view.
Acknowledgements
We thank Reza Shadmehr and Paul Cisek for their comments on an earlier
version of this paper. This work was supported in part by a Center for
Excellence grant (8006/00) administered by the Israel Science Foundation
(ISF), by the Bundesministerium für Bildung und Forschung – DeutschIsraelische Projektkooperation(BMBF – DIP), by grant 2001073 administrated by the Binational Science Foundation (BSF), and by a special
contribution of the Golden Charitable Trust. R.P. was supported by a
Constantiner fellowship.
References
1 Marr, D. (1982) Vision: A Computational Investigation into the Human
Representation and Processing of Visual Information, W. H. Freeman
2 Gilbert, C.D. et al. (2001) The neural basis of perceptual learning.
Neuron 31, 681 – 697
502
Review
TRENDS in Neurosciences Vol.27 No.8 August 2004
3 Ahissar, M. and Hochstein, S. (1997) Task difficulty and the specificity
of perceptual learning. Nature 387, 401 – 406
4 Pouget, A. and Snyder, L.H. (2000) Computational approaches to
sensorimotor transformations. Nat. Neurosci. 3 (Suppl.), 1192 – 1198
5 Buneo, C.A. et al. (2002) Direct visuomotor transformations for
reaching. Nature 416, 632 – 636
6 Schaal, S. and Atkeson, C.G. (1998) Constructive incremental learning
from only local information. Neural Comput. 10, 2047– 2084
7 Donchin, O. et al. (2003) Quantifying generalization from trial-by-trial
behavior of adaptive systems that learn with basis functions: theory
and experiments in human motor control. J. Neurosci. 23, 9032– 9045
8 Chou, I.H. and Lisberger, S.G. (2002) Spatial generalization of
learning in smooth pursuit eye movements: implications for the
coordinate frame and sites of learning. J. Neurosci. 22, 4728 – 4739
9 Thoroughman, K.A. and Shadmehr, R. (2000) Learning of action
through adaptive combination of motor primitives. Nature 407,
742 – 747
10 Coltz, J.D. et al. (1999) Cerebellar Purkinje cell simple spike discharge
encodes movement velocity in primates during visuomotor arm
tracking. J. Neurosci. 19, 1782– 1803
11 Dragoi, V. et al. (2000) Adaptation-induced plasticity of orientation
tuning in adult visual cortex. Neuron 28, 287– 298
12 Godde, B. et al. (2002) Plasticity of orientation preference maps in
the visual cortex of adult cats. Proc. Natl. Acad. Sci. U. S. A. 99,
6352 – 6357
13 Schuett, S. et al. (2001) Pairing-induced changes of orientation maps
in cat visual cortex. Neuron 32, 325 – 337
14 Weinberger, N.M. (1998) Physiological memory in primary auditory
cortex: characteristics and mechanisms. Neurobiol. Learn. Mem. 70,
226 – 251
15 Suga, N. et al. (2002) Plasticity and corticofugal modulation for
hearing in adult animals. Neuron 36, 9 – 18
16 Chen, L.L. and Wise, S.P. (1996) Evolution of directional preferences
in the supplementary eye field during acquisition of conditional
oculomotor associations. J. Neurosci. 16, 3067– 3081
17 Wise, S.P. et al. (1998) Changes in motor cortical activity during
visuomotor adaptation. Exp. Brain Res. 121, 285 – 299
18 Taylor, D.M. et al. (2002) Direct cortical control of 3D neuroprosthetic
devices. Science 296, 1829 – 1832
19 Carmena, J.M. et al. (2003) Learning to control a brain – machine
interface for reaching and grasping by primates. PLoS Biol. 1, E42
20 Li, C.S. et al. (2001) Neuronal correlates of motor performance and
motor learning in the primary motor cortex of monkeys adapting to an
external force field. Neuron 30, 593 – 607
21 Padoa-Schioppa, C. et al. (2002) Neuronal correlates of kinematics-todynamics transformation in the supplementary motor area. Neuron
36, 751 – 765
22 Paz, R. et al. (2003) Preparatory activity in motor cortex reflects
learning of local visuomotor skills. Nat. Neurosci. 6, 882– 890
23 Gribble, P.L. and Scott, S.H. (2002) Overlap of internal models in motor
cortex for mechanical loads during reaching. Nature 417, 938 – 941
24 Classen, J. et al. (1998) Rapid plasticity of human cortical movement
representation induced by practice. J. Neurophysiol. 79, 1117– 1123
25 Dragoi, V. et al. (2002) Dynamics of neuronal sensitivity in visual
cortex and local feature discrimination. Nat. Neurosci. 5, 883 – 891
26 Sakai, M. and Suga, N. (2001) Plasticity of the cochleotopic (frequency)
map in specialized and nonspecialized auditory cortices. Proc. Natl.
Acad. Sci. U. S. A. 98, 3507 – 3512
27 Krakauer, J.W. et al. (2000) Learning of visuomotor transformations
for vectorial planning of reaching trajectories. J. Neurosci. 20,
8916 – 8924
28 Snippe, H.P. (1996) Parameter extraction from population codes:
a critical assessment. Neural Comput. 8, 511 – 529
29 Pouget, A. et al. (2003) Inference and computation with population
codes. Annu. Rev. Neurosci. 26, 381 – 410
30 Zhang, K. and Sejnowski, T.J. (1999) Neuronal tuning: to sharpen or
broaden? Neural Comput. 11, 75– 84
31 Muller, J.R. et al. (1999) Rapid adaptation in visual cortex to the
structure of images. Science 285, 1405 – 1408
32 Chowdhury, S.A. and Suga, N. (2000) Reorganization of the frequency
map of the auditory cortex evoked by cortical electrical stimulation
in the big brown bat. J. Neurophysiol. 83, 1856– 1863
www.sciencedirect.com
33 Recanzone, G.H. et al. (1993) Plasticity in the frequency representation of primary auditory cortex following discrimination training
in adult owl monkeys. J. Neurosci. 13, 87 – 103
34 Schoups, A. et al. (2001) Practising orientation identification improves
orientation coding in V1 neurons. Nature 412, 549 – 553
35 Chen, L.L. and Wise, S.P. (1997) Conditional oculomotor learning:
population vectors in the supplementary eye field. J. Neurophysiol. 78,
1166– 1169
36 Laubach, M. et al. (2000) Cortical ensemble activity increasingly
predicts behaviour outcomes during learning of a motor task. Nature
405, 567 – 571
37 Sanger, T.D. (2003) Neural population codes. Curr. Opin. Neurobiol.
13, 238 – 249
38 Paz, R. and Vaadia, E. (2004) Learning-induced improvement in
encoding and decoding of specific movement directions by neurons
in the primary motor cortex. PLoS Biol. 2, E45
39 Sommer, M.A. and Wurtz, R.H. (2002) A pathway in primate brain for
internal monitoring of movements. Science 296, 1480 – 1482
40 Kawato, M. (1999) Internal models for motor control and trajectory
planning. Curr. Opin. Neurobiol. 9, 718– 727
41 Flanagan, J.R. et al. (2003) Prediction precedes control in motor
learning. Curr. Biol. 13, 146 – 150
42 Gaffan, D. (1996) Associative and perceptual learning and the concept
of memory systems. Cognit. Brain Res. 5, 69– 80
43 Lee, T.S. et al. (2002) Neural activity in early visual cortex reflects
behavioral experience and higher-order perceptual saliency. Nat.
Neurosci. 5, 589 – 597
44 Bar-Yosef, O. et al. (2002) Responses of neurons in cat primary
auditory cortex to bird chirps: effects of temporal and spectral context.
J. Neurosci. 22, 8619– 8632
45 Machens, C.K. et al. (2004) Linearity of cortical receptive fields
measured with natural sounds. J. Neurosci. 24, 1089– 1100
46 Donchin, O. et al. (1998) Primary motor cortex is involved in bimanual
coordination. Nature 395, 274– 278
47 Steinberg, O. et al. (2002) Neuronal populations in primary motor
cortex encode bimanual arm movements. Eur. J. Neurosci. 15,
1371– 1380
48 Semple, M.N. and Kitzes, L.M. (1993) Binaural processing of sound
pressure level in cat primary auditory cortex: evidence for a
representation based on absolute levels rather than interaural level
differences. J. Neurophysiol. 69, 449– 461
49 Crist, R.E. et al. (1997) Perceptual learning of spatial localization:
specificity for orientation, position, and context. J. Neurophysiol. 78,
2889– 2894
50 Crist, R.E. et al. (2001) Learning to see: experience and attention in
primary visual cortex. Nat. Neurosci. 4, 519 – 525
51 Ulanovsky, N. et al. (2003) Processing of low-probability sounds by
cortical neurons. Nat. Neurosci. 6, 391 – 398
52 Wolpert, D.M. and Ghahramani, Z. (2000) Computational principles
of movement neuroscience. Nat. Neurosci. 3 (Suppl.), 1212– 1217
53 Cohn, J.V. et al. (2000) Reaching during virtual rotation: context
specific compensations for expected coriolis forces. J. Neurophysiol. 83,
3230– 3240
54 Wolpert, D.M. and Kawato, M. (1998) Multiple paired forward and
inverse models for motor control. Neural Netw. 11, 1317 – 1329
55 Haruno, M. et al. (2001) Mosaic model for sensorimotor learning and
control. Neural Comput. 13, 2201 – 2220
56 Shadmehr, R. and Brashers-Krug, T. (1997) Functional stages in the
formation of human long-term motor memory. J. Neurosci. 17,
409– 419
57 Krakauer, J.W. et al. (1999) Independent learning of internal models
for kinematic and dynamic control of reaching. Nat. Neurosci. 2,
1026– 1031
58 Tong, C. et al. (2002) Kinematics and dynamics are not represented
independently in motor working memory: evidence from an interference study. J. Neurosci. 22, 1108 – 1113
59 Flanagan, J.R. et al. (1999) Composition and decomposition of internal
models in motor learning under altered kinematic and dynamic
environments. J. Neurosci. 19, RC34
60 Tong, C. and Flanagan, J.R. (2003) Task-specific internal models for
kinematic transformations. J. Neurophysiol. 90, 578– 585
61 Gandolfo, F. et al. (1996) Motor learning by field approximation. Proc.
Natl. Acad. Sci. U. S. A. 93, 3843 – 3846
Review
TRENDS in Neurosciences Vol.27 No.8 August 2004
62 Karniel, A. and Mussa-Ivaldi, F.A. (2002) Does the motor control
system use multiple models and context switching to cope with a
variable environment? Exp. Brain Res. 143, 520 – 524
63 Osu, R. et al. (2004) Random presentation enables subjects
to adapt to two opposing forces on the hand. Nat. Neurosci. 7,
111 – 112
503
64 Hepp-Reymond, M.C. et al. (1999) Context-dependent force
coding in motor and premotor cortical areas. Exp. Brain Res. 128,
123– 133
65 Martin, J.H. (1989) Neuroanatomy: Text and Atlas, 2nd edn,
Elsevier
3rd INMED/TINS conference
THE MULTIPLE FACETS OF GABAERGIC SYNAPSES
Sept. 15–18, 2004. Théâtre du Golfe, La Clotat – France
Organised by Y. BEN-ARI
Information on: www.inmed.univ-mrs.fr
BEN-ARI Yehezkel (Marseille – France)
GABA, brain maturation and epilepsies
BERNARD Christophe (Marseille – France)
Fate of GABAergic inhibition during epileptogenesis
BUSZAKI György (Newark – USA)
Grouping of cell assemblies by oscillation in the hippocampus
CHERUBINI Enrico (Trieste – Italy)
Coincidence detection enhances GABA release at mossy-fibreCA3 synapses in the developing hippocampus
COSSART Rosa (France & USA)
Imaging oscillators in brain structures
FREUND Tamas (Budapest – Hungary)
Presynaptic control of GABAergic transmission by
endocannabinoids in the cerebral cortex
GOZLAN Henri (Marseille – France)
Developmental aspects of endo-canabinoid transmission
GUTIERREZ Rafael (Mexico)
The dual glutamatergic–GABAergic phenotype of the
hippocampal granule cells
HESTRIN Shaul (Stanford – USA)
Networks of GABAergic neurons in the neocortex
HOUANG Josh Z (New York – USA)
Molecular mechanisms underlying the subcellular targeting of
GABAergic synapses
KAILA Kai (Helsinki – Finland)
Two developmental switches in GABAergic signaling:
KCC2 and CA VII
KANDLER Karl (Pittsburgh – USA)
Developmental refinement of an inhibitory sound localization
circuit
LLINAS Rodolfo (New York – USA)
Cognition as an intrinsic resonant brain state that simulates
reality
LIU Guosong (Cambridge MA – USA)
Organization of excitatory and inhibitory synaptic connections
on dendritic tree and its implications for neural computation
MARTY Alain (Paris – France)
Short term plasticity of GABAergic synapses
McBAIN Chris (Bethesda – USA)
Novel mechanisms of plasticity at hippocampal mossy fiber
interneuron synapses
MILES Richard (Paris – France)
Interneurons and population activities in the hippocampus
MODY Istavan (Los Angeles – USA)
Tonic inhibition and epilepsy
MONYER Hannah (Heidelberg – Germany)
Genetic manipulations to study sunchronous network activity
RAKIC Pasco (New Haven – USA)
Development and evolution of GABAergic interneurons
SERNAGOR Evelyne (Newcastle upon Tyne –UK)
Control of retinal maturation by changes in GABAergic
function and by visual experience
SOLTESZ Ivan (California – USA)
Activity-dependent potentiation of endocannabinoid
signaling at GABAergic synapses
SPITZER Nicholas C (La Jolla – USA)
Neurotransmitter specification: decisions and revisions that
minutes can reverse
KHAZIPOV Roustem (Marseille – France)
Networks of GABAergic neurons in the neocortex
STERIADE Mircea (Quebec – Canada)
Role of thalamic reticular GABAergic neurons in normal and
paroxysmal oscillations
KULLMANN Dimitri (London – UK)
Excitatory GABA in the early cortical activities
KLAUSBERGER Thomas and SOMOGYI Peter (Oxford – UK)
Space and time in hippocampal GABAergic connections
www.sciencedirect.com