Download Descending Systems Translate Transient Cortical Commands into a

Cerebral Cortex August 2012;22:1904– 1914
doi:10.1093/cercor/bhr267
Advance Access publication September 30, 2011
Descending Systems Translate Transient Cortical Commands into a Sustained Muscle
Activation Signal
Uri Shalit1, Nofya Zinger1, Mati Joshua2 and Yifat Prut1
1
Department of Medical Neurobiology and The Interdisciplinary Center for Neural Computation, The Hebrew
University—Hadassah Medical School, Jerusalem 91120, Israel and 2Department of Physiology, W.M. Keck Foundation Center for
Integrative Neuroscience, University of California, San Francisco, San Francisco, CA 94143, USA
Shalit and Zinger have contributed equally to this work
Address correspondence to Yifat Prut, Department of Medical Neurobiology, The Hebrew University—Hadassah Medical School, PO Box 12272,
Jerusalem 91120, Israel. Email: [email protected].
Controlling motor actions requires online adjustments of timevarying parameters. Although numerous studies have attempted to
identify the parameters coded in different motor sites, the
relationships between the temporal profile of neuronal responses
and the dynamics of motor behavior remain poorly understood in
particular because motor parameters such as force and movement
direction often change over time. We studied time-dependent
coding of cortical and spinal neurons in primates performing an
isometric wrist task with an active hold period, which made it
possible to segregate motor behavior into its phasic and sustained
components. Here, we show that cortical neurons transiently code
motor-related parameters when actively acquiring a goal, whereas
spinal interneurons provide persistent information regarding maintained torque level and posture. Moreover, motor cortical neurons
differed substantially from spinal neurons with regard to the
evolvement of parameter-specific coding over the course of a trial.
These results suggest that the motor cortex and spinal cord use
different control policies: Cortical neurons produce transient motor
commands governing ensuing actions, whereas spinal neurons
exhibit sustained coding of ongoing motor states. Hence, motor
structures downstream to M1 need to integrate cortical commands
to produce state-dependent spinal firing.
Keywords: motor control, motor cortex, mutual information, spinal cord,
tuning
Introduction
During movement, the motor system is required to continuously
adjust the machinery which produces actions while integrating
general information about the environment (e.g., spatial location
of targets) and the internal state (e.g., limb configuration).
Activity in the motor cortex has been shown to be related to
both low-level (Kalaska et al. 1989, 1997; Cabel et al. 2001;
Kurtzer et al. 2005, 2006) and high-level (Georgopoulos et al.
1982, 1986; Schwartz et al. 2004) parameters but the coding of
these attributes has often been found to be interdependent
(Scott and Kalaska 1995), suggesting that these 2 representations may coexist in the activity of single motor cortical neurons
(Kalaska 2009). Clearly at some stage, this mixed representation
must be resolved since motoneurons (MNs) produce an
unequivocal muscle command. Several mechanisms have been
proposed to explain the way muscle commands are extracted
from the cortical representation. In one scenario, a small group
of motor cortical neurons specifically translates the mixed
cortical representation into a muscle-based command, which is
subsequently transmitted downstream to muscles (Kakei et al.
Ó The Author 2011. Published by Oxford University Press. All rights reserved.
For permissions, please e-mail: [email protected]
1999, 2001). Alternatively, the extensive connectivity of the
motor cortex with multiple subcortical sites (Baldissera et al.
1981; He et al. 1993; Porter and Lemon 1993; Lemon 2008)
could be used to extract the muscle command from the
integrated cortical representation (Shah et al. 2004).
Recently, based on the finding that intrinsic commands
emerge first downstream to M1, we hypothesized that spinal
interneurons do not serve as a simple integrator of cortical
signals but act as a processing link in the translation of motor
commands into muscle-based signals (Yanai et al. 2007, 2008;
Harel et al. 2008). To further test this hypothesis, we studied
the response properties of neurons in the motor cortex, spinal
cord, and of working muscles, while 3 monkeys performed
a 2D wrist task designed to dissociate motor parameters (Kakei
et al. 1999).
Here, we show that the directional-torque tuning of single
cortical and spinal neurons had a similar shape but different
spatiotemporal distributions: Cortical neurons expressed a temporally transient and spatially uniform preferred direction (PD)
distribution consistent with a brief representation of the target
space, whereas spinal PDs were spatially biased but temporally
stable, consistent with the muscle activation profile. Analyzing
the information content of cortical and spinal neurons further
distinguished the transient nature of cortical coding from the
enhanced sensitivity of spinal neurons to static motor states.
These results are in line with a distributed mode of processing in
the motor system in which the motor cortex codes motor
commands which integrate both the required ‘‘muscle-based’’
parameters of the action and the externally defined target to be
obtained, while the spinal circuitry maintains the ongoing motor
state. These distinct coding properties indicate that elements
downstream to M1 further process motor commands to make it
suitable for driving working muscles.
Materials and Methods
Animals and Behavioral Task
Three female monkeys (Maccaca fascicularis, 3--4 kg) performed an
isometric 2D wrist task with an instructed delay period (Fig. 1a).
During task performance, the monkey held its hand in either
a pronation or supination position and controlled a cursor on
a computer screen by applying a 2D isometric torque at the wrist
(flexion/extension and radial/ulnar). A trial was initiated by the
appearance of a central target (Trial onset). The monkey positioned
the cursor inside the target by generating zero torque for a rest period
(500--600 ms). Then, 8 uniformly distributed peripheral targets
appeared around the center target at a fixed distance, defining the
onset of a delay period. One of these peripheral targets was presented
in a distinct color for 500 ms (Cue). The disappearance of the central
Figure 1. Experimental setup. (a) Monkeys were seated in front of a computer screen and controlled an onscreen cursor via wrist torque. During task performance, we recorded
the activity of spinal units, cortical units, and relevant muscles. Radial torque and behavioral events (the timing of which is marked by green vertical lines) recorded in a single trial
are shown (for details, see Materials and Methods section). The trial is aligned on torque onset (time zero) and a small arrow indicates the beginning of torque offset. Also shown
are the activities of cortical and spinal neurons recorded simultaneously in this trial. Note that during the precue period, the spinal firing rate was higher than the cortical firing
rate. Also, at torque onset the cortical neuron exhibited a phasic response (red trace) while the spinal neurons (blue trace) exhibited a tonic response. These properties were
characteristic of our dataset. (b) Location of motor cortical recording sites obtained for the 3 monkeys (D, G, and V) and spinal sites obtained for 2 of these monkeys (D and V).
Maps were overlaid manually to align the locations of central and arcuate sulci. The cortical map depicts low-threshold arm-related sites in which tuned neurons were found
(squares) and sites which did not meet the above criteria but included cells with at least 50 trials in both postures (circles).
target (monkey V, 1300--1700 ms, monkeys D and G, 850--1200 ms after
target onset) served as the ‘‘Go’’ signal. The monkey then had to shift
the cursor into the previously cued target by generating an isometric
torque ramp (Torque onset) in the appropriate direction and
magnitude and keep the cursor within the target for an active hold
period (350--750 ms, monkeys G and V, 700--1000 monkey D). The
reappearance of the central target (occurring simultaneously with the
disappearance of the peripheral targets) signaled the end of the active
hold period. The monkey returned to the rest position (Torque offset)
and received a reward, after which the screen went blank for 1000-1500 ms and a new trial started. The timing between externally induced
events (e.g., between Cue onset and the Go signal) was randomly varied
from trial to trial. The required torque level during the hold period was
adjusted for each monkey to obtain persistent muscle activity without
causing fatigue. During the hold period, torque values were in the
range of 30 mNM (mean, 28.6 mNM, median, 27.36 mNM) with some
variation across monkeys (median of 33 mNM, 25 mNM, and 27 mNM
for monkeys D, G, and V, respectively). The allowed torque variation
once the cursor was in the target was less than 5 mNM (5.9, 3.88, and
4.96 mNM for monkeys D, G, and V, respectively). Times of torque
onset and offset were defined offline as crossing of a threshold by the
derivative of the torque signal. These events (the beginning of Torque
onset and the beginning of Torque offset) were used to align neural
activity during analysis of neuron responses.
Recording Sessions
Details of the recording technique are described elsewhere (Yanai et al.
2007). A cortical chamber was implanted above the motor cortex, and
the location of the arm-related primary motor cortex (M1) was mapped
using a train of stimulating pulses (50 ms of biphasic stimulation
delivered at 300 Hz with intensity <60 lA). In this study, we selected
cortical neurons that were recorded from the arm-related area in the
primary motor cortex (threshold for evoked response <15 lA).
Subsequently, a spinal chamber was implanted above the cervical
spinal cord (C6--T1). Extracellular single-unit activity was recorded
simultaneously from the motor cortex and spinal interneurons (INs)
located at intermediate lamina while the monkey performed the task
(using 1--2 cortical electrodes and a single spinal electrode). We
targeted spinal neurons that were located deep in the intermediate to
ventral spinal lamina recorded deeper than 1 mm from the first cells
encountered. These neurons differed in firing properties from dorsal
horn neurons as well as from MN (Prut and Perlmutter 2003a, 2003b).
We did not encounter any MN action potentials when averaging muscle
activity around spike firing times, although a small proportion of false
negatives could have been possible.
We recorded electromyogram (EMG) activity using either subcutaneous (monkey V) or transcutaneous (monkey D and G) electrodes
from selected identified arm and forearm muscles. These included the
Flexor-carpi-ulnaris (FCU), Palmaris Longus (PL), Flexor-carpi-radialis
(FCR), Flexor-digitorum-superficialis (FDS), Flexor-digitorum-profundos
(FDP), Pronator-Teres (PT), Extensor-carpi-ulnaris (ECU), Extensordigitorum-carpi (EDC), Extensor-carpi-radialis (ECR), Extensor-digitorum-45 (ED45), Extensor-digitorum-23 (ED23), Abductor-pollici-longus
(APL) Biceps, Triceps, and Deltoid. After the recordings were completed,
2 monkeys (D and G) were deeply anaesthetized (ketamine, pentobarbital, 30 mg/kg), and pins were inserted into known coordinates of the
cortical implant. The animals were then euthanized with pentobarbital
sodium (50 mg/kg iv) and perfused with 10% formalin or 4%
paraformaldehyde. Cortical locations of penetrations relative to anatomical landmarks were subsequently reconstructed (Fig. 1b).
Data Analysis
Response Properties of Neurons
To test the tuning properties of cells, we selected cells for which we
had more than 40 correct trials per posture and at least 6 targets for
which we had more than 5 trials per target.
We first tested cells for significant PD (resampling method
[Crammond and Kalaska 1996], P < 0.05) during the –200 to +500 ms
window around torque onset. Here, we compared the length of the
Cerebral Cortex August 2012, V 22 N 8 1905
Table 1
Recorded muscles þ recorded muscles that were tuned and had a unimodal PD—recorded
muscles that were not used for analysis
Table 2
Numbers of tuned cortical and spinal cells
Tuned cells and muscles
Extensors
APL
EDC
ECR
ECU
ED23
ED45
Monkey D
Monkey G
Monkey V
Flexors
Monkey D
Monkey G
Monkey V
Proximals
Monkey D
Monkey G
Monkey V
þ
þ
-þ
FDS
þ
þ
-Triceps
-þ
þ
þ
þ
FCU
þ
þ
-Deltoid
þ
--FDP
þ
þ
þ
þ
þ
þ
þ
þ
PT
þ
þ
-FCR
þ
þ
þ
Biceps
þ
þ
PL
þ
þ
þ
Cellsa
Pro datab
Sup datab
213 (64, 95, 54)
83 (34, 0, 49)
43 (15, 21, 7)
157
54
38
128
48
30
158 (44, 78, 36)
52 (20, 0, 32)
109
26
103
34
Conditionale
Unconditionalf (monkey D)
274
151
109
15
Data for MI analysis
þ
Motor cortex
Spinal cord
vector obtained by vector summation of the directional activity of the
cell with the expected length obtained after random assignment of
direction of the single trial data and recomputation of the length of the
summed activity. The shuffling method was repeated 4000 times to
allow testing of the statistical significance of the result. In order to
further study tuning properties, we looked for cells in which tuning
was unimodal by testing their tuning for significant fit to a von Mises
function (Amirikian and Georgopoulos 2000) (Matlab function ‘‘lsqcurvefit’’) at a significance level of P < 0.01. The von Mises distribution is
the circular analog of the normal distribution on a line. The function we
fitted had 4 parameters:
f x a; b; h; j =a + b e jcosðx – hÞ ;
where a represents baseline rate, b modulation, j tuning width, and h
PD. To compute the von Mises fit, we considered cells with a number of
trials exceeding the above criteria for a minimal number of trials. The
significance test was based on the explained variance of the fit, and only
cells with a significant PD and a significant VM fit (F-test, P < 0.01) were
considered to be tuned cells.
Computing the tuning properties during torque offset was carried
out in a similar manner but during a time window of –200 to +500 ms
around the time at which the monkey started to shift the cursor back to
the central target (i.e., the end of the active hold period).
Perievent Time Histogram
For cells which passed the above criteria, we computed the perievent
time histogram (PETH) around torque onset and torque offset as
defined above. In both cases, the aligning event was the beginning of
torque production, which was defined offline based on the derivative of
the torque signal (Yanai et al. 2007). The PETH was obtained by
counting spikes in 20-ms bins. We then averaged these histograms for
the target nearest the neuron’s PD across all neurons (spinal and
cortical neurons separately). This procedure yielded a population-based
temporal profile of activity in the PD.
For data analysis of muscle activity, we considered only those
muscles that expressed significant unimodal tuning (Table 1). Averaged
muscle response around torque onset was normalized by first removing
the baseline and dividing by the standard deviation, both computed
during a test epoch of –1000 to –500 ms relative to torque onset. These
response profiles were divided by their norm to obtain a comparable
range of values across the different muscles.
The response profile of single neurons and muscles was quantified
based on the average response around torque onset in the PD. Average
rates were calculated within a 0--250 ms time window after torque onset
(fr1) as well as during a late window spanning 250 ms before torque
offset (fr2) using a PETH aligned on that event. A ‘‘phasic-tonic index’’
(PTI) was then defined as (fr1 – fr2)/(fr1 + fr2). This index tends toward
zero when the response profile is tonic (i.e., fr1 fr2) and toward 1
when the response profile is phasic (fr1 > > fr2). For EMG activity, fr1 was
computed for a later time window (150--400 ms) because the muscle
response lagged behind neural activity. We found that the exact timing
of the test window (fr1) selected for EMG data did not have a noticeable
impact on the distribution width, but its median monotonically increased
from slightly negative values to near zero values, reflecting the fact that
early windows captured the rising phase of the EMG signal.
1906 Translation of Transient Cortical Commands into a Sustained Signal
Torque onsetc
Primary motor cortexd
Spinal
Muscles
Torque offset
Primary motor cortexd
Spinal
d
Shalit et al.
319 (109, 0, 210)
166 (15, 0, 151)
a
Counts the total number of cells, with the contribution of monkey D, G, and V in parentheses.
A single cell could have been tested inpronation and\or supination trials, so that N(all calls) $
N(pro_data) þ N(sup_data).
c
For analysis of tuning and PETH shape, we took cells that were tuned around torque onset (or
offset) based on a resampling method (P \ 0.05) and had a significant fit to the Von Mises
function (P \ 0.01).
d
Arm-related areas of M1.
e
For computing MI, we took cells from motor cortex and spinal cord with at least 50 trials in
pronation and supination. Note that only monkey V had spinal neurons which met this criterion.
f
To compute the unconditional MI for monkey D, we studied cells with at least 50 trials in either
hand posture. In this case, single cells could have contributed values for both postures.
b
Computing Mutual Information for Neuronal Data
Mutual information (MI) (Cover and Thomas 2006) between F, the
number of spikes in a given window, and torque direction D, was
calculated by using the joint probability p(f, d) of a cell firing f spikes in
the window, while the torque direction was d. The overall probability
p(f) of observing
f spikes in the window was obtained by margin
alization:p f =+d pðf ; dÞ; p(d) was calculated likewise.
Given these probabilities, the MI (in bits) between F and D was
defined to be:
pðf ; dÞ
:
ð1Þ
I F ; D = + p f ; d log2
pðf Þ pðdÞ
f ;d
In practice, the probabilities p(f,d) are approximated by the empirical
probabilities p̂ðf ; dÞ for every cell and given time window: given
a count matrix C whose (i, j) entry is the number of times j spikes were
observed for torque direction i, we have p̂ðf =j; d =iÞ=+Cði;jÞ
:
i;j Cði;jÞ
This leads to an empirical approximation of the MI:
p̂ðf ; dÞ
;
ð2Þ
Iˆ F ; D = + p̂ f ; d log2
p̂ðf Þ p̂ðdÞ
f ;d
where p̂ðf Þ and p̂ðdÞ are derived by marginalization of p̂ðf ; dÞ.
As there were 8 different torque directions and a typical cell can emit
15 spikes within a 500-ms window, there were as many as 8 3 15 = 120
different probabilities to assess, with about 100 or so trials conducted.
This resulted in undersampling, which in turn is known to lead to
a positive bias in the estimation of MI values (Treves and Panzeri 1995;
Panzeri et al. 1996).
In the current work, we used a method developed by Nelken et al.
(2005), which has been shown to consistently and relatively robustly
offset this positive bias. The method is as follows.
Treves and Panzeri (1995) calculated an asymptotic expansion of the
expected positive bias encountered when calculating MI based on
empirical probabilities derived from count matrices. The expansion is
in inverse powers of the number of samples N =+i; j Cði; jÞ. For an m 3
k count matrix with N samples, the first term in this expansion is
ðm – 1Þðk – 1Þ
ð3Þ
2 ln2 N
(the other terms depend on the hypothetical true underlying
probability and thus cannot be calculated).
Nelken et al.’s (2005) ‘‘Adaptive Direct’’ algorithm exploits this bias
estimate to find an optimum binning strategy for the count matrix C. At
each step, the ‘‘naı̈ve’’ MI is calculated as in equation (2), and the
approximate bias is subtracted following equation (3). This gives an
estimate of the true MI. After this is calculated, the matrix C is reduced
by adjoining the least probable column (or row) to its least probable
neighbor. In our setup, columns correspond to the number of spikes
emitted; for example, if a cell occasionally fired 8 spikes and 6 spikes,
but never 7 spikes, the 7 spike column would have a probability of
0 and would be merged with one of its 2 neighbors (the one with the
lower probability). Typically, as the 8 torque directions were more or
less evenly spread, whereas the different spike numbers were far from
being so, initially the algorithm merged columns until they were all
roughly equiprobable with the rows. The merging ends when the
matrix has one row or column.
With each binning step, the naı̈ve MI is reduced but so is the bias.
The binning in which the difference between the 2 is greatest is
considered optimal from a debiasing point of view. The bias-corrected
MI value obtained by this binning method was the MI value used in
further analyses. In order to further prevent overestimation of MI, only
cells with at least 50 trials per posture were included in these analyses.
Note that these cells were not necessarily directionally tuned. We used
2 window sizes for the analyses, spanning either 500 or 300 ms.
Window sliding was in steps of 10 ms.
MI conditioned on posture, I(F ;DjP) was calculated in a similar way:
pðf ; djposÞ
I F ; D P = + pðposÞ + p f ; d pos log2
pðf jposÞ pðdjposÞ
pos
f ;d
= + pðposÞI F ; D posture=pos ;
pos
where p(pos) is the probability of each of the different postures
(pronation or supination). I(F ;PjD) was calculated analogously.
Data Sets Used for Analysis
We recorded a total of 555 motor cortical cells (178, 103, and 274 from
monkeys D, G, and V, respectively) and 292 spinal cells (66, 0, and 226).
To analyze the tuning and response properties of spinal and cortical
neurons (Table 2), we considered tuned motor cortical and spinal
neurons (more than 40 trials per posture and at least 6 targets for
which the number of trials >5) with a significant fit to the von Mises
function. For this purpose, motor cortical neurons were selected from
sites that evoked an arm response when applying stimulation at an
intensity <15 lA. To compute the MI, we used all the motor cortical
and spinal cells with at least 50 trials in pronation and supination from
monkeys D and V for which we had both cortical and spinal data. In this
case, we used a more lenient criterion for including motor cortical
neurons that were not necessarily from low-threshold and\or armrelated sites. To study the effect of extended delay (used in monkey D
only), we used the MI between firing rate and direction given a single
hand posture—I(D ; FjP = p).
Table 2 summarizes the number of cells and muscles used for the
different analyses.
Results
Differential Dynamic of Response Patterns Observed for
Cortical Neurons Versus Spinal Neurons
The task we used required the monkeys to shift the cursor to
a precued peripheral target and to generate a persistent and
constant torque level to maintain the cursor within the
peripheral target until a release signal was given, at which
point the monkey shifted the cursor back to center (Fig. 1a).
We studied the evolution of cortical and spinal activity in the
PD for tuned neurons. We found that in the PD, cortical
neurons often expressed a transient response, which was more
pronounced during torque ramp but decayed during active
hold (Fig. 2a). On the other hand, INs fired persistently
throughout the torque ramp and active hold periods in the PD.
At the end of the hold period, the sustained activity of tuned
Figure 2. Response profiles of spinal and cortical neurons. PETHs were computed for
tuned cortical (red) and spinal (blue) neurons. PETHs were computed for the PD in
a time window spanning 500 to þ1000 ms around torque onset and using a 20-ms
bin size. For each event, we show the average response across all neurons (upper
chart) and the average normalized response obtained after scaling the single-cell
response to vary between 0 and 1 (lower chart) so that the effect of firing rate
variation across cells was canceled out. The number of cortical and spinal PETHs used
for this analysis is shown in the upper 2 panels (Nctx and Nsp, respectively). (a)
PETHs computed around torque onset for the target nearest the neuronal PD. (b)
Same as a but using the torque offset event for aligning the PETHs. In this case, tuned
cells and PD were computed for the period of 200 to 500 ms around torque offset.
The number of neurons is thus different from that in a.
INs was reduced and returned to baseline levels (Fig. 2b). By
contrast, tuned cortical neurons were recruited during this
period and thus responded in a transient manner. For
comparison, we computed the response profile of active
muscles with significant unimodal tuning properties (Fig. 3a).
We found that the average response profile of muscles
(distal and proximal) tended to be tonic, similar to spinal INs
(Fig. 3b).
We further quantified single cell response profiles using an
ad-hoc formulated PTI, which was based on the firing at onset
and the sustained hold period (Fig. 4a). The index was the
difference between the 2 values over their sum. Note that on
average, the onset of muscle activity lagged behind the onset of
neuronal activity. For this reason, we used different time
windows to compute the PTI for neural and muscle activity
(see Materials and Methods). Near-zero values corresponded to
tonic response and values near 1 corresponded to the phasic
response profile. We found (Fig. 4b) that spinal neurons tended
to concentrate around zero values, whereas the mean of the
cortical distribution was shifted toward positive values. The
difference between the 2 distributions was significant (t-test,
P < < 0.001). The PTI for muscles was narrowly distributed
around zero, suggesting a high tendency for muscles to fire
tonically in our task. This result was consistent for a variety of
time windows used for computing the muscle PTI.
We further tested and found that motor cortical neurons that
were either recorded from sites in which cortical stimulations
evoked a short latency spinal response (Yanai et al. 2007) or had
a functional link with forearm muscles (measured by spikeCerebral Cortex August 2012, V 22 N 8 1907
Figure 3. Response profiles of forearm muscles. Activity of forearm muscles was recorded trans- or subcutaneously. (a) The average muscle activity around torque onset was
computed for each of the 8 targets. Examples of extensor muscle (extensor carpi radialis-ECR) in pronation trials are shown together with the tuning curves and corresponding PD.
The average radial torque is shown for each target (dashed lines). Average muscle activity in the target nearest the PD is highlighted in black. (b) The response of flexors,
extensors, and proximal muscles in the PD was averaged separately. Each plot shows the overlay of a single muscle response profile (light gray) and the average profile (black).
This figure shows the result obtained for pronation trials.
Figure 4. Quantification of single cells and muscle response profiles. (a) Average response profiles of cortical neurons (red), spinal neurons (blue), and working muscles (green).
Standard error of the mean is shown around each profile. As neurons and muscles have a different scale, the profiles were adjusted arbitrarily to allow comparison of the
response shape rather than response magnitude. Time windows for which average response was measured for computing the PTI are shown for neurons (fr2 and fr1) and muscles
(fr2 and fr1’). Note that fr2 was computed using a PETH that was aligned on torque offset and is shown here solely for purposes of illustration. (b) Distribution of PTIs (see text for
formal definition) computed for cortical neurons (orange), spinal neurons (blue), and muscles (green). Arrows point to the mean level for each distribution. (c) Cumulative
distribution of the PTIs shown in b. Lines (50% and 95%) are shown. (d) Distributions of PTIs computed for cortical neurons recorded from sites connected with spinal sites (red),
unconnected cortical sites (orange), and cortical sites for which we found a functional linkage with forearm muscles (brown).
1908 Translation of Transient Cortical Commands into a Sustained Signal
d
Shalit et al.
triggered averaging) had a similar PTI as did unidentified motor
cortical neurons (Fig. 4d, Kruskal Wallis test, P > 0.65).
Finally, to test whether the same group of neurons were
activated at onset and offset of active torque, we computed the
preferred target difference for cortical and spinal neurons that
were significantly tuned on both torque onset and torque offset
(Fig. 5a,b). We found that the average difference in preferred
target was around 180° for cortical neurons, indicating that
these cells were activated for 2 opposing targets during torque
onset and offset (Fig. 5c). In contrast, spinal PD differences
were around 0°, indicating that cells were activated for the
same targets during torque onset and offset (Fig. 5d). Hence, if
we consider the sustained active-hold period as a single motor
state, this state is maintained by a single spinal population,
whereas distinct groups of cortical neurons mediate its onset
and offset (in our case, shifting the cursor to target and back to
center). Examples of single cortical and spinal neurons that
expressed this differential target-related activity are shown in
Figure 6a. Note that here we computed the preferred target
and not the preferred torque direction, as used more often.
Nevertheless, the difference in tuning consistency between
cortical and spinal neurons was maintained even when
examining the preferred torque direction.
Cortical and Spinal Neurons Have a Similar Tuning
Shape but a Different Distribution of PDs
The PETH captures the response pattern of neurons for a specific
(preferred) target, whereas the tuning function of a neuron
captures its mean response across different targets. Tuning
Figure 5. PD differences between torque onset and offset. We selected neurons,
which were tuned both at onset and offset of torque and plotted the distributions of
single neuron PD differences (dPDs) between these 2 epochs. The polar histograms
show that for cortical neurons (a) the dPDs tended to concentrate around 180°,
whereas for spinal neurons (b), these values were closer to zero. Numbers in gray
next to each polar histogram denote the scale (in counts) for each plot. (c and d)
Illustrations showing for cortical and spinal neurons with the same PD during toque
onset (target 2—dark gray) what would be their preferred target at torque offset. For
convenience, targets are numbered from 1 to 8. Torque onset and offset are denoted
with dashed lines and arrows. Note that for cortical neurons (c), maximal firing during
torque offset would be when monkeys shift the cursor from the opposite target
(target 6) back to center. In contrast, spinal neurons (d) with the same PD at torque
onset are expected to be have maximal firing rate during torque offset when monkeys
shift the cursor from the preferred target (target 2) back to center.
properties are commonly used to quantify the relationships
between neuronal firing and external or internal parameters (Fig.
6a,b). Tuning functions of cortical and spinal neurons were similar
for all parameters except for the baseline level that was higher for
INs (Fig. 6c). Specifically, the distribution of tuning width
(quantified as the width at half height) was nonsignificantly
different for cortical and spinal neurons (Wilcoxon test for equal
medians, P > 0.85). This result indicates that downstream motor
structures do not operate by ‘‘sharpening’’ tuning functions.
Despite the similar tuning function of single spinal and
cortical neurons, their respective PD distributions were
considerably different. Cortical neurons expressed uniformly
distributed PDs, whereas the PDs of INs were significantly biased
(Fig. 6d, Rao test, P < 0.05). The observed bias rotated clockwise
with hand posture, as expected from the distribution of muscle
PDs (Fig. 6d). In pronation trials, this bias was consistent with an
overrepresentation of flexor muscles, as was shown for spinal INs
(Perlmutter et al. 1998). Interestingly, for supination trials, the
downward bias remained with an added recruitment of neurons
with PDs in lateral and upward directions. This added activation
in supination trials might be related to the fact that performing
the trials in this posture was more demanding and hence
required a broader activation of INs. Nevertheless, although
cortical PDs were consistent with the uniform distribution of
targets in the torque\external space, spinal PDs bore more
resemblance to the nonuniform PDs of working muscles.
MI between Neuronal Firing and Motor Parameters
Temporal response patterns and neuronal tuning are not
necessarily independent properties of neuronal coding.
Changes in response profile in the PD can be accompanied
by rate modulation occurring for other (nonpreferred)
directions such that directionality is either enhanced or
canceled out. To investigate the possible interdependency
between these 2 properties, we computed the MI between
a cell’s firing rate and the coded variables (target direction and
hand posture). MI expresses the reduction of uncertainty that
a neuron’s activity provides about the value of a specific
parameter, so that the MI between neurons located at different
sites and a specific parameter can be compared directly.
Intuitively, for a specific neuronal population, when the
average MI (measured in bits) between a neuron and
a parameter is high, a smaller set of independent neurons are
required to fully encode this parameter. In our paradigm, 3 bits
were required to code the 8 targets and one bit was required to
code the 2 postures. We measured the conditional MI of
cortical and spinal firing in relation to either target direction
(MID) or hand posture (MIP). The average MID around torque
onset (Fig. 7a) revealed only a slight tendency of cortical
neurons to be more informative while firing at lower rates
compared with INs. Nonetheless, the time-resolved MID
revealed 2 unique patterns: cortical neurons were significantly
more informative before and during torque onset compared
with INs (Fig. 7b), but during the late hold period, this trend
was reversed and the spinal MID was higher, as can be
appreciated when aligning the time-dependent MID on torque
offset (Fig. 7c, plotted on expanded time and information axes).
This behavior was particularly visible in trials with a longer
active hold period (Fig. 7d) for which the dynamics of cortical
MID was unaffected by the extended delay, while the spinal MID
faithfully followed the torque profile. Note that when the hold
Cerebral Cortex August 2012, V 22 N 8 1909
Figure 6. Response properties of cortical and spinal neurons. (a) Raster plots of cortical (left) and spinal (right) units. Trials were aligned on torque onset and sorted according to
target direction. The 700-ms epoch around torque onset used for computing the tuning shown in panel b is marked with a colored bar above each raster plot. (b) Single trial
average firing rate (computed for a time window of 200 to þ500 ms around torque onset) and the von Mises fit of the same cortical and spinal neurons shown in a. Preferred
targets and number of trials (n) are shown for each unit. (c) The average tuning curves (obtained from cells with significant PDs and significant fit to the von Mises function) were
computed for cortical (red) and spinal (blue) neurons. The figure shows (from left to right) the average tuning, the average tuning after removing the baseline from the single cell
tuning function to reflect differences in the tuning gain, and the normalized tuning (after baseline removal) to reflect differences in tuning width. (d) Distribution of PDs for cortical
neurons (left), spinal neurons (middle), and working muscles (right) in pronation (upper charts) and supination (lower charts) trials. A Rao test was used to determine significant
deviation from uniform distribution and the obtained P values (RAOp) are shown. For muscles, distribution of PDs for flexors (red), extensors (blue), and proximal muscles (green),
computed for the same time window, are shown separately.
period is short (relative to the bin size used in the analysis), the
difference between cortical and spinal MID is more difficult to
capture since it is ‘‘diluted’’ by direction-related cortical
information that reappears at torque offset. During this period,
1910 Translation of Transient Cortical Commands into a Sustained Signal
d
Shalit et al.
spinal MID decreases while cortical MID increases so that the
difference between these 2 measures becomes smaller. Due to
the bin size used in this analysis, this trend affected the
temporal profile of MID even before torque offset.
Figure 7. MI between neuronal firing and target direction. (a) Relationships between average firing rate (across all targets) and average target-related information in a given
hand posture computed for cortical (red) and spinal (blue) neurons. Values were computed for a time window spanning 200 to þ500 ms around torque onset. Each neuron
contributed either 1 or 2 data points depending on the number of trials available in each hand posture. Large dots represent the average across each population. (b) Time-resolved
conditional MI between neuronal firing and target direction (MID), computed around torque onset using overlapping time windows of 500 ms, shifted 10 ms each time step. (c)
MID aligned on torque offset in the ordinate and computed for time windows preceding torque offset (i.e., active hold). In this case, we used 300-ms time windows to maintain
temporal resolution despite the shorter duration of the active hold period. Both x-axis and y-axis are expanded in this plot. (d) Same as b but using data from one monkey only
(monkey D) where the hold period was longer. In this case, we took spinal cells with at least 50 trials in either pronation or supination but not both (see Materials and Methods).
(e) conditional MI between neuronal firing and hand posture (MIP) computed around torque onset for each target and then averaged across targets.
In this study, task performance required a correct estimate
of hand posture that was constant within trials but changed
across trials (either pronation or supination). Analysis of posturerelated MI revealed that hand posture was encoded differently by
the cortical and spinal populations: spinal MIP was high
throughout trial performance and consistently higher than
cortical MIP (Fig. 7e). At torque onset, cortical MIP momentarily
reached the level of spinal MIP. This result is congruent with the
above result showing that spinal neurons are more informative
about static motor states, whereas cortical neurons provide
transient (yet often superior) information around times of
change in the motor state.
Discussion
Early studies of the motor system formulated the question of
cortical motor control in a categorical manner that contrasted
a high level (Georgopoulos et al. 1982; Georgopoulos 1996;
Moran and Schwartz 1999) with a low level (Kalaska et al. 1997;
Todorov 2000; Scott et al. 2001; Morrow et al. 2007) coding
scheme. More recent studies of the system have shown that
motor cortical activity cannot be fully explained by either of
these models alone (Scott and Kalaska 1997; Sergio and Kalaska
1997; Yanai et al. 2008; Kalaska 2009). Our study further
explored the neuronal control of voluntary movements by using
2 different approaches: first, a direct comparison of cortical and
spinal neurons was used to highlight the evolvement of coding at
different levels of the motor system. Second, an analysis of timedependent changes in coding explored the consistency of
coding during persistent motor states. In this manner, we
showed first of all that the cortical response profile is transient,
whereas spinal neurons fire persistently during torque production. Second, we found that while cortical neurons express
Cerebral Cortex August 2012, V 22 N 8 1911
a uniform distribution of PDs, spinal PDs were biased. Finally,
studying MI between neuronal firing and task parameters
revealed that cortical neurons are more sensitive to timechanging parameters while spinal neurons provide superior
coding for static parameters. The discrepancy in both parametric
coding properties and response profiles indicates that spinal
activity cannot be modeled as a simple averaging of cortical
command; rather, a more sophisticated decoding mechanism is
required to translate the cortical command into the observed
spinal activity. The consistency between spinal activity and
muscle activity further indicates that elements downstream to
M1 are more likely to contain a muscle-based motor command.
Cortical neurons exerted a predominantly transient command that contained a corepresentation of both intrinsic
variables (hand posture) and extrinsic parameters (spatial
location of targets). This activity preceded the motor action
and thus had predictive value regarding the ensuing action.
However, as soon as the target was reached, cortical activity
waned substantially, despite the sustained torque produced by
the monkey. Thus, the cortical command may not be able to
function adequately as a low-level control signal because of its
temporal properties which deviate from the exerted torque
profile and its parametric content (i.e., a combined representation of internal and external parameters) as we showed
previously when studying the coordinate frame of cortical
neurons in the same task (Yanai et al. 2008).
The activity of spinal neurons is much more appropriate to
function as a low-level muscle controller. The data here show
that the temporal properties of this signal and its parametric
content are closer to the actual state of the motor system at any
given point. These results suggest that spinal neurons not only
operate in the same coordinate frame as muscles (Yanai et al.
2008) but they provide a sustained drive to muscles when such
torque profile is required. Our findings thus further extend
earlier reports (Crammond and Kalaska 1996; Johnson et al.
1996) by suggesting that from the premotor cortex down to
the spinal cord, there is a shift in response pattern (from phasic
to tonic) and response timing (from premovement to sustained
movement). Nevertheless, further verification is needed given
the paradigmatic differences between our work and previous
studies (in terms of task, working muscles, and data analysis).
The task used in this study may not provide a realistic
picture of normal motor behavior. Motor actions are often
transient and contain no sustained period of static torque. In
these cases, we expect to obtain similar temporal pattern of
cortical, spinal, and muscle activity. However, this similar
activation pattern does not necessarily mean similar coding. In
this respect, our task can be viewed not as a way of replicating
normal motor conditions but rather as a way to dissociate the
coding properties of cortical and spinal neurons.
Earlier studies described the distribution of PDs of cortical
neurons as uniform (Georgopoulos et al. 1982), biased in
a manner consistent with motor habits (Naselaris et al. 2006),
or bimodal in a manner similar to the distribution of muscle
PDs (Scott et al. 2001; Kurtzer et al. 2006; Herter et al. 2007). In
our study, we limited behavior to a wrist task and found
a discrepancy in the behavior of cortical and spinal neurons:
Although cortical neurons had uniformly distributed PDs, spinal
PDs were biased in a manner which at least in some cases was
consistent with an uneven representation of flexor versus
extensor muscles (Perlmutter et al. 1998). This discrepancy in
PD distribution between cortical and spinal neurons might be
1912 Translation of Transient Cortical Commands into a Sustained Signal
d
Shalit et al.
interpreted as further evidence for the transformation of the
more general cortical command into a muscle-based command
(Harel et al. 2008).
In this study, we employed a population-based approach,
pooling single neurons in each site to reveal a global coding
scheme. This approach is consistent with previous attempts to
identify motor cortical signals (Georgopoulos et al. 1986).
Clearly, averaging across all neurons may mask the heterogeneity of the single-cell response profile. Nonetheless, analyzing
single cell (and muscle) response profiles confirmed the
validity of this approach as we found no multimodal distribution of response properties as might be expected if there were
a few subpopulations among cortical (or spinal) neurons.
Moreover, we found no unique relations between subgroups of
neurons (e.g., cortical neurons recorded from CS sites in which
stimulation evoked an orthodromic spinal response) and
response profiles. This indicates that the phasic response
profile of cortical neurons is a general phenomenon that also
holds for output sites, suggesting that the translation to a tonic
pattern most likely occurs downstream to M1. This result
appears to be at odds with some evidence suggesting that the
motor cortex contains specific subgroups of neurons which
may act as low-level controllers, such as corticomotoneuronal
(CM) cells which in some cases fire in a manner consistent with
the temporal profile of the produced torque (Fetz and Cheney
1979; Cheney and Fetz 1980). It should be noted, however, that
CM neurons constitute only a small fraction of motor cortical
neurons (Porter and Lemon 1993), especially in terms of wrist
movements and even these neurons have diverse response
profiles (Hepp-Reymond et al. 1978). In addition, for spinal INs,
no correlation was found between response profile and
connectivity with muscles (Perlmutter et al. 1998), thus
making it reasonable to expect that the response profile of
neurons should be more tightly correlated with their input
pattern and not with their output connectivity.
It should also be noted here that many studies have
emphasized the diversity of spinal INs (Jankowska 1992),
a feature which was not addressed in our study. It may have
been the case that our criterion for using unimodally tuned INs
to measure response profiles restricted the population of
tested INs to a more homogeneous subgroup, as reflected by
the narrow distribution of PTIs. A more comprehensive analysis
is required to further identify these neurons in terms of their
input and output connectivity.
The transient nature of the cortical signal could suggest
cortical control of time-derivative motor parameters (e.g.,
velocity or torque derivative in isometric conditions) acting
only at times of change in torque levels (Moran and Schwartz
1999). Furthermore, the discrepancy in preferred targets found
for cortical neurons during torque onset and offset may suggest
that cortical neurons code for torque direction, whereas spinal
neurons code for torque level. However, we found that cortical
information about target location substantially preceded the
actual action and hence was available hundreds of milliseconds
before any torque change could be detected (or controlled for)
by this system. This delay is longer than the estimated lag
between cortical signal and muscle activity, which was
reported to be about 100 ms (Moran and Schwartz 1999).
Furthermore, the possibility that cortical neurons simply code
for torque derivative is inconsistent with our previous results
(Yanai et al. 2008) showing that motor cortical neurons
operate in a different coordinate frame than muscles.
Therefore, while motor cortical neurons often code information at times of change in the motor state, they do not appear
to code the time derivative of low-level parameters per se.
Combining the results obtained in this study with previous
results showing that just before movements cortical neurons
operate in a coordinate frame which is in-between external and
internal suggests that to translate the cortical signal into a muscle
command the following steps take place: 1) rotation and scaling of
arriving signals to complete the transformation from target space
into muscle space and 2) transformation of a transient command
into sustained firing which is suitable for continuous activation of
the muscles. This process is responsible for decoding the
integrated cortical command to eliminate the nonrelevant
space-based components it contains. The exact subcortical site
(or sites) at which these events take place remains unclear, and
the involvement of subcortical but supraspinal sites cannot be
ruled out. However, several spinal mechanisms may contribute to
these processes. Specific termination patterns of corticospinal
fibers may generate an independent representation of some motor
parameters while averaging out redundant (e.g., external-related)
information. Intrasegmental circuitry may act as a ‘‘latch’’ in
translating a transient signal into sustained firing and further
enhancing an independent representation of motor variables. In
this context, neuromodulatory control (Kiehn et al. 1996; Heckman
et al. 2004) may participate in prolonging cortical impact. Lastly,
interactions between peripheral and spinal circuitry may provide
the required rescaling of the uniform distribution of PDs into
a biased, muscle-like representation as well as contributing to the
persistent coding of stable motor states such as posture.
Funding
Israel Science Foundation grants (ISF-1490/09); Binational
Science Foundation (BSF-2007442); Baruch Foundation; Rosetrees Foundation.
Notes
We thank Eli Nelken for insightful discussions which greatly
contributed to this work. Conflict of Interest : None declared.
References
Amirikian B, Georgopoulos AP. 2000. Directional tuning profiles of
motor cortical cells. Neurosci Res. 36:73--79.
Baldissera F, Hultborn H, Illert M. 1981. Integration in spinal neuronal
systems, V.B. Brooks, Editor, Handbook of Physiology. The Nervous
System, Vol. II Am. Physiol. Soc, Bethesda pp. 509--595 Motor control.
Part 1.
Cabel DW, Cisek P, Scott SH. 2001. Neural activity in primary motor
cortex related to mechanical loads applied to the shoulder and
elbow during a postural task. J Neurophysiol. 86:2102--2108.
Cheney PD, Fetz EE. 1980. Functional classes of primate corticomotoneuronal cells and their relation to active force. J Neurophysiol.
44:773--791.
Cover TM, Thomas JA. 2006. Elements of information theory. New York:
Wiley-Interscience.
Crammond DJ, Kalaska JF. 1996. Differential relation of discharge in
primary motor cortex and premotor cortex to movements versus
actively maintained postures during a reaching task. Exp Brain Res.
108:45--61.
Fetz EE, Cheney PD. 1979. Muscle fields and response properties
of primate corticomotoneuronal cells. Prog Brain Res. 50:
137--146.
Georgopoulos AP. 1996. On the translation of directional motor cortical
commands to activation of muscles via spinal interneuronal systems.
Brain Res Cogn Brain Res. 3:151--155.
Georgopoulos AP, Kalaska JF, Caminiti R, Massey JT. 1982. On the
relations between the direction of two-dimensional arm movements
and cell discharge in primate motor cortex. J Neurosci. 2:1527--1537.
Georgopoulos AP, Schwartz AB, Kettner RE. 1986. Neuronal population
coding of movement direction. Science. 233:1416--1419.
Harel R, Asher I, Cohen O, Israel Z, Shalit U, Yanai Y, Zinger N, Prut Y.
2008. Computation in spinal circuitry: lessons from behaving
primates. Behav Brain Res. 194:119--128.
He SQ, Dum RP, Strick PL. 1993. Topographic organization of
corticospinal projections from the frontal lobe: motor areas on
the lateral surface of the hemisphere. J Neurosci. 13:952--980.
Heckman CJ, Kuo JJ, Johnson MD. 2004. Synaptic integration in
motoneurons with hyper-excitable dendrites. Can J Physiol Pharmacol. 82:549--555.
Hepp-Reymond MC, Wyss UR, Anner R. 1978. Neuronal coding of static
force in the primate motor cortex. J Physiol (Paris). 74:287--291.
Herter TM, Kurtzer I, Cabel DW, Haunts KA, Scott SH. 2007.
Characterization of torque-related activity in primary motor cortex
during a multijoint postural task. J Neurophysiol. 97:2887--2899.
Jankowska E. 1992. Interneuronal relay in spinal pathways from
proprioceptors. Prog Neurobiol. 38:335--378.
Johnson PB, Ferraina S, Bianchi L, Caminiti R. 1996. Cortical networks
for visual reaching: physiological and anatomical organization of
frontal and parietal lobe arm regions. Cereb Cortex. 6:102--119.
Kakei S, Hoffman DS, Strick PL. 1999. Muscle and movement representations in the primary motor cortex. Science. 285:2136--2139.
Kakei S, Hoffman DS, Strick PL. 2001. Direction of action is represented
in the ventral premotor cortex. Nat Neurosci. 4:1020--1025.
Kalaska JF. 2009. From intention to action: motor cortex and the
control of reaching movements. Adv Exp Med Biol. 629:139--178.
Kalaska JF, Cohen DA, Hyde ML, Prud’homme M. 1989. A comparison of
movement direction-related versus load direction-related activity in
primate motor cortex, using a two-dimensional reaching task. J
Neurosci. 9:2080--2102.
Kalaska JF, Scott SH, Cisek P, Sergio LE. 1997. Cortical control of
reaching movements. Curr Opin Neurobiol. 7:849--859.
Kiehn O, Johnson BR, Raastad M. 1996. Plateau properties in
mammalian spinal interneurons during transmitter-induced locomotor activity. Neuroscience. 75:263--273.
Kurtzer I, Herter TM, Scott SH. 2005. Random change in cortical load
representation suggests distinct control of posture and movement.
Nat Neurosci. 8:498--504.
Kurtzer I, Herter TM, Scott SH. 2006. Nonuniform distribution of reachrelated and torque-related activity in upper arm muscles and
neurons of primary motor cortex. J Neurophysiol. 96:3220--3230.
Lemon RN. 2008. Descending pathways in motor control. Annu Rev
Neurosci. 31:195--218.
Moran DW, Schwartz AB. 1999. Motor cortical representation of speed
and direction during reaching. J Neurophysiol. 82:2676--2692.
Morrow MM, Jordan LR, Miller LE. 2007. Direct comparison of the taskdependent discharge of M1 in hand space and muscle space. J
Neurophysiol. 97:1786--1798.
Naselaris T, Merchant H, Amirikian B, Georgopoulos AP. 2006. Largescale organization of preferred directions in the motor cortex. I.
Motor cortical hyperacuity for forward reaching. J Neurophysiol.
96:3231--3236.
Nelken I, Chechik G, Mrsic-Flogel TD, King AJ, Schnupp JW. 2005.
Encoding stimulus information by spike numbers and mean
response time in primary auditory cortex. J Comput Neurosci.
19:199--221.
Panzeri S, Biella G, Rolls ET, Skaggs WE, Treves A. 1996. Speed, noise,
information and the graded nature of neuronal responses. Network.
7:365--370.
Perlmutter SI, Maier MA, Fetz EE. 1998. Activity of spinal interneurons
and their effects on forearm muscles during voluntary wrist
movements in the monkey. J Neurophysiol. 80:2475--2494.
Porter R, Lemon RN. 1993. Corticospinal function and voluntary
movement. Oxford: Calrendon Press.
Cerebral Cortex August 2012, V 22 N 8 1913
Prut Y, Perlmutter SI. 2003a. Firing properties of spinal interneurons
during voluntary movement. I. State-dependent regularity of firing. J
Neurosci. 23:9600--9610.
Prut Y, Perlmutter SI. 2003b. Firing properties of spinal interneurons
during voluntary movement. II. Interactions between spinal
neurons. J Neurosci. 23:9611--9619.
Schwartz AB, Moran DW, Reina GA. 2004. Differential representation of
perception and action in the frontal cortex. Science. 303:380--383.
Scott SH, Gribble PL, Graham KM, Cabel DW. 2001. Dissociation
between hand motion and population vectors from neural activity in
motor cortex. Nature. 413:161--165.
Scott SH, Kalaska JF. 1995. Changes in motor cortex activity during
reaching movements with similar hand paths but different arm
postures. J Neurophysiol. 73:2563--2567.
Scott SH, Kalaska JF. 1997. Reaching movements with similar hand
paths but different arm orientations. I. Activity of individual cells in
motor cortex. J Neurophysiol. 77:826--852.
1914 Translation of Transient Cortical Commands into a Sustained Signal
d
Shalit et al.
Sergio LE, Kalaska JF. 1997. Systematic changes in directional
tuning of motor cortex cell activity with hand location in
the workspace during generation of static isometric forces
in constant spatial directions. J Neurophysiol. 78:1170-1174.
Shah A, Fagg AH, Barto AG. 2004. Cortical involvement in the
recruitment of wrist muscles. J Neurophysiol. 91:2445--2456.
Todorov E. 2000. Direct cortical control of muscle activation in
voluntary arm movements: a model. Nat Neurosci. 3:391--398.
Treves A, Panzeri S. 1995. The upward bias in measures of
information derived from limited data samples. Neural Comput.
7:399--407.
Yanai Y, Adamit N, Harel R, Israel Z, Prut Y. 2007. Connected
corticospinal sites show enhanced tuning similarity at the onset of
voluntary action. J Neurosci. 27:12349--12357.
Yanai Y, Adamit N, Israel Z, Harel R, Prut Y. 2008. Coordinate
transformation is first completed downstream of primary motor
cortex. J Neurosci. 28:1728--1732.