Unraveling adaptation and mutual inhibition in
Functional Neurobiology and Helmholtz Institute,
Utrecht University, Utrecht, The Netherlands
Martin J. M. Lankheet
When the visual system is confronted with incompatible images in the same part of the visual field, the conscious percept
switches back and forth between the rivaling stimuli. Such spontaneous flips provide important clues to the neuronal basis
for visual awareness. The general idea is that two representations compete for dominance in a process of mutual inhibition,
in which adaptation shifts the balance to and fro. The inherent nonlinear nature of the rivalrous flip-flop and its stochastic
behavior, however, made it impossible to disentangle inhibition and adaptation. Here we report a general method to
measure the time course, and asymmetries, of mechanisms involved in perceptual rivalry. Supported by model simulations,
we show the dynamics of opponent interactions between mutual inhibition and adaptation. The findings not only provide
new insight into the mechanism underlying rivalry but also offer new opportunities to study and compare a wide range of
bistable processes in the brain and their relation to visual awareness.
Keywords: adaptation, binocular rivalry, dynamics, mutual inhibition, reverse correlation
Perceptual rivalry behaves like a flip-flop in which
percepts remain stable for a while and then abruptly
switch from one percept to the other (Wheatstone, 1838).
Changes in cortical processing associated with such
spontaneous alternations have been studied extensively
because they may expose the neuronal basis for visual
awareness (Crick & Koch, 1998). Despite many neuro-
physiological (Leopold & Logothetis, 1996; Logothetis,
1998; Logothetis & Shall, 1989) and fMRI (Lee & Blake,
2002; Polonsky, Blake, Braun, & Heeger, 2000; Tong,
Nakayama, Vaughan, & Kanwisher, 1998) studies aimed
at identifying neuronal activity engaged in rivalry, the
mechanisms involved remain puzzling (Blake & Logothetis,
One of the main reasons is that the rivalrous flop-flop is
highly nonlinear and stochastic. Noise in activity levels
associated with both percepts causes a random distribu-
tion of dominance intervals (Levelt, 1965). In line with
the notion of a nonlinear flip-flop is the finding that
interactions between rivalrous percepts are strongly
asymmetrical (Levelt, 1965). Increasing the contrast in
one eye may have a large effect on the duration of the
percept associated with the other eye, whereas it may
hardly affect durations for the higher contrast stimulus.
Since Levelt formulated his second law, this asymmetry
has been a hallmark in explanations of rivalry processes.
Rivalry models, such as the two-stage model proposed by
Wilson (2003) and Laing and Chow’s (2002) spiking
neural model, pass this critical test.
The asymmetries arise from the interplay between at
least two dynamic processes. On the one hand there is
mutual inhibition, operating on a relatively short time
scale. On the other hand there is adaptation, operating on
a larger time scale. Both mechanisms are activity depen-
dent but in fundamentally different ways. High levels of
activity favor inhibition of the opposite perceptual
representation. This might shorten dominance intervals
for the suppressed eye and increase durations for the
dominant eye. Adaptation works the other way around:
high activity levels cause stronger adaptation and may
therefore decrease durations for the dominant eye.
Because these two processes act on different time scales,
activity changes at different moments during perceptual
dominance intervals are likely to affect rivalry in different
ways. Thus, to understand rivalry it is important to know
the time course of rivalrous interactions.
So far, distributions of interval durations have been
the main source of information on dynamics of rivalry,
but clearly this does not allow one to separate
dynamics of inhibition and adaptation. One obvious
approach to solve this problem would be to present
probes of stimulus variations at different points in time
during the flip cycle (Norman, Norman, & Bilotta,
2000). This has several disadvantages: First, timing of
flashes relative to the next flip is variable and cannot be
controlled directly. Second, to study possible mutual
interactions, multiple combinations of probes in one
stimulus and the other and at different strengths must be
used. Such measurements would take a long time,
especially because single dominance durations normally
last seconds already.
Instead of probes, we used a reverse correlation
technique, in which signal strength in both stimuli is
modulated according to a random sequence. By correlat-
ing perceptual flips to signal strengths, we recover the full
time course of rivalrous interactions. Moreover, we can
pinpoint any asymmetries in the process by comparing
Journal of Vision (2006) 6, 304–310 http://journalofvision.org/6/4/1/304
doi: 10.1167/6.4.1 Received June 29, 2005; published February 21, 2006ISSN 1534-7362 * ARVO
correlations for the two stimuli and the two directions of
flips. The reverse correlation technique described here
provides the necessary extra handle to disentangle effects
of adaptation, inhibition, and noise and thus sheds a new
light on previous and future results.
Here we present results for binocular motion rivalry
(Alais & Blake, 1998; Blake, Yu, Lokey, & Norman,
1998; Meng, Chen, & Qian, 2004; van de Grind, van Hof,
van der Smagt, & Verstraten, 2001). Random dot patterns
for the left and for the right eye were presented side by
side on a 19-in. CRT display, running at 100 Hz.
Observers viewed the stimuli with the help of a mirror
stereoscope in a dimly lit room. Great care was taken to
properly adjust the viewing apertures and vergence angle
for each observer to create a single dichoptic image that
was easily fixated. Proper alignment of left and right eye
was supported by a central fixation mark and by the
contours of the patterns. Each random dot pattern
consisted of 6000 bright dots on a dark background,
displayed in a 2.2 degree square window (200 ? 200
pixels), resulting in 15% dot density. Stimuli were
generated in real time and were updated on every frame.
The coherent motion was to the upper left in one eye and
to the upper right in the other, at a velocity of 1.6 deg/s
(1 pixel/frame). Dot positions were kept as floats and
rounded towards the nearest monitor pixel for drawing.
This minimizes the effect of step size differences in
horizontal/vertical directions and diagonal directions for
the rectangular pixel array. Orthogonal motion directions
in the two eyes were chosen to minimize Bmonocular[
The difference in motion directions in the two eyes
caused clear binocular rivalry. Observers reported domi-
nance of one motion direction or the other, with clear flips
in between. Piece-meal rivalry, in which different parts of
the display seem to move in different directions, was
mostly limited to a wave of transition across the display.
Transitions generally occurred fast compared to percep-
To measure the time course of activity changes
affecting dominance flips, we modulated the coherence
of the motion around a clearly supra-threshold value. The
coherence value is defined as the percentage of dots
taking part in the coherent motion while the remaining
dots moved in a random direction. In area MT of
macaques, which is strongly related to motion perception,
variations in supra-threshold coherence values cause
nearly linear variations in cell-activity (Britten, Shadlen,
Newsome, & Movshon, 1992). We assumed that motion
detectors in our brains that are involved in binocular
motion rivalry were equally well modulated by motion
coherence. The noncoherent dots of the pattern moved
with the same step size, but in a direction drawn from a
random distribution for each dot, on each frame. This type
of noise causes minimal segregation of motion and noise.
To further minimize visual segregation of stimulus and
noise dots, dots were given a limited lifetime of 300 ms.
Dots were refreshed asynchronously on every frame.
Temporal modulations of motion coherence do not
result in clearly visible transients. On a slower time scale,
observers see variations in coherence, but fast variations
do not reach visual awareness. As a result, modulations do
affect the process of rivalry but do not inevitably force
flips. Coherence values were drawn from a uniform noise
distribution on each frame (frame rate 100 Hz) and
subsequently filtered with a first order low-pass filter
(time constant 500 ms), which resulted in Gaussian white
noise. Amplitude and time constant of coherence varia-
tions were chosen to obtain sufficient resolution in both
time and coherence domains. Temporal filtering assured
that stimulus energy was confined to an effective range of
temporal frequencies. Results described here were
obtained with a mean coherence value of 50% and
standard deviations of the modulations (after filtering) of
All observers had ample experience in rivalry experi-
ments and in other motion detection experiments.
Monocular motion coherence thresholds were below 5%
for each observer. All observers, except MLA, were naive
regarding the outcome of the experiment. They were
instructed to indicate their perceptual flips by pressing
arrow keys on the keyboard. The stimuli were generated
in real time and were shown for as long as required to
finish a session. Flips were collected in three or four
sessions of about 20 min, with long periods in between.
Observers were also free to rest at any time during a
session. The first three flips after a break were ignored, so
as to rule out start-up transients. Coherence modulations
and key presses were stored at a resolution of one frame
for off-line analysis.
Figure 1 shows an example of noise modulations for
both stimuli and explains the analysis procedure. Flips
from left to right and from right to left were used to
calculate peri-flip motion strength correlation functions.
These correlations describe sensitivity at each point in
time preceding flips for coherence changes in one eye and
in the other eye. This yields four correlation functions
(Figure 2): two for averaged modulations while the right
eye percept was dominant and two for left eye dominance.
To avoid confounding effects of previous flips, we
averaged for one complete interval only, plus a period
of 2 s following flips. Figure 2b shows the cumulative
distribution of interval durations. Mean interval duration
for observer MLA was about 5 s. The number of averaged
noise traces thus declines rapidly for increasing interval
duration. As a result, the data become less reliable for
Journal of Vision (2006) 6, 304–310Lankheet305
Results for five observers are summarized in Figure 3.
The first column shows the distribution of interval
durations. Observers differ in the parameters of the
distribution; both in mean interval duration and in the
shape of the distribution. Especially observer MLO flips
more frequently than the other subjects, a finding
consistent with results in other rivalry tasks. The lack of
long intervals for this observer renders the correlations at
times before about j3 s less reliable. Whereas the
parameters of the interval distributions are the main
source of information in other studies and have been
analyzed in great detail, we only focus on the reverse
Because the two ipsilateral curves (noise for the
dominant stimulus) and the two contralateral curves
(suppressed stimuli) were very similar, we averaged the
two. The second column in Figure 3 shows the averaged
coherence values for noise modulations in the suppressed
and dominant eye. The most surprising finding is that the
correlation functions show a biphasic profile. Directly
preceding a flip, there is a period of about 1 s in which
increments in the suppressed eye and decrements in the
dominant eye favor the occurrence of flips. At longer
delays, about 1Y6 s before flips, modulations have the
opposite effect. This opposite effect is smaller in
amplitude but lasts considerably longer. Despite the large
differences in dominance interval durations, all observers
show qualitatively the same result.
Effects for the dominant eye are opposite to those for
the suppressed eye. Closer examination, however, shows
that the effects are not mirror images. Typically, the short
latency phase is more pronounced for the suppressed
eye than for the dominant eye. An increment in the
suppressed eye is more effective than a decrement in the
Modulation amplitudes differ between observers. Inter-
estingly, the modulation amplitude seems to correlate
positively with the mean interval duration. Observer MLO
flips at a high frequency and shows a small modulation
amplitude, whereas RB has the longest dominance
durations and shows the highest modulation amplitude.
MLA, JD, and RvW are intermediate, both in the
amplitudes of the modulation functions and in the value
of the mean interval duration.
The difference in modulation amplitudes for the domi-
nant and suppressed stimuli covaries with the overall
amplitude of the modulation functions. It is especially
strong for observers showing a relatively small effect. For
Figure 1. Time course of motion coherence values for left and
right eye. A new coherence value was drawn from a uniform
distribution on each frame (every 10 ms) and then passed
through a single first order low-pass filter with a time constant of
500 ms. This resulted in a Gaussian distribution of coherence
values with a standard deviation of 11.5% around the mean of
50%. The triangular symbols indicate the observers’ reports of
flips from left to rightward motion and vice versa. The red and
blue interrupted lines show the resulting dominance intervals.
Figure 2. Peri-flip correlation functions (a), constructed by
averaging the noise traces corresponding to separate dominance
intervals, as indicated by the interrupted lines in Figure 1. This
results in four flip-triggered averages: two for noise in the
dominant stimulus (red and black curves) and two for noise in
the suppressed stimulus (green and blue). (b) The number of
intervals used in averaging each 10-ms time bin. The number
decreases for longer intervals relative to the flip, and hence the
variability in the averages increases. Data shown are for observer
Journal of Vision (2006) 6, 304–310Lankheet306
observer MLO, the negative modulation for the dominant
effect, this asymmetry between dominant and suppressed
stimuli is much smaller. The effects for dominant and
suppressed stimuli not only differ in amplitude but also
have clearly different time courses. Sensitivity for varia-
tions in the dominant eye change gradually during the flip
cycle, whereas the function for the suppressed eye shows a
fairly abrupt change from a negative to a positive value.
These results indicate that rivalry is a complex
interaction between (at least) two processes, with differ-
ent time constants. As a result, flip-triggered noise
averages show a biphasic profile. At short latencies
relative to flips, the effect is consistent with mutual
inhibition, at longer latencies the sign of the effect
reverses and is consistent with an adaptation effect. This
seems counterintuitive at first but makes sense consider-
ing differences in dynamics between adaptation and
mutual inhibition. To gain further insight in the inter-
actions between adaptation and mutual inhibition under-
lying the flip-triggered averages, we performed a simple
model study in which we simulate the experiments
The sole aim of the model simulation was to show that
a minimal model already shows the behavior we
observed, indicating that it reflects a key property of
rivalry mechanisms. Other constraints will surely require
more sophisticated models but that is of no concern here.
Figure 4 shows a diagram of the rivalry model, which is
a simplified version of previously proposed rivalry models
(Lehky, 1988; Matsuoka, 1984; Wilson, 2003). It consists
of two processing streams (or Bneurons[), one for the left
eye and one for the right eye. Each neuron receives the
coherence level as input. In the model, processing for left
and right eye is similar and to a high degree independent,
except for the mutual negative feedback. Signals in each
stream pass adaptation, a filter stage, and a compression
stage. A Weber-type of adaptation is implemented by
dividing the input by the sum of fixed constant, c, and a
low-pass filtered version of the input. The time constant
for adaptation, Cad, determines to a large extent the mean
duration of dominance intervals. It was set to 1 s. To
account for low temporal resolution and long temporal
Figure 3. Flip-triggered noise averages for different observers.
The left-hand column shows the distributions of interval durations
for left (red) and right eye (blue). Data were collected for 250Y450
leftward and rightward flips. The right-hand column shows results
for stimuli in the suppressed eye (black) and dominant eye
(green), with coherence values on the left-hand axis. The thin red
and blue lines show the number of intervals used in averaging
(right-hand axis). Confidence intervals correspond to T2 standard
errors of the mean.
Figure 4. Model diagram. The model consists of two streams, one
for the right and one for the left eye. The output provides negative
feedback to the contra lateral side. The input (stimulus plus
feedback) passes an adaptation stage in which the input is
divided by a low-pass filtered (time constant Cad) version of the
input (plus a fixed constant C). The adapted signal is low-pass
filtered (time constant Ci) and passed through a nonlinear,
NakaYRushton type of compression stage. A "homunculus"
compares left and right eye outputs and reports a flip whenever
the balance of responses reverses.
Journal of Vision (2006) 6, 304–310Lankheet307
integration times in motion coherence detection, we
includedalow-passfilterwithatimeconstantCiof 500 ms.
The low-pass filtered activity level is converted into a
final output by a nonlinear, NakaYRushton type of
compression stage (x ¼ xn=xnþ c), where c (set to 0.5)
determines the position and n (set to 2) determines the
width of the operating range. The final comparison stage
reports a flip whenever the balance of responses reverses.
The feedback is characterized by a gain factor (set to 96).
An additional filter stage was included in the feedback
path but its time constant, Cfb, was short (20 ms) relative
to the other time constants and therefore less influential.
The model contains no intrinsic noise sources, and no
response latency between flip detection and Bbutton
press[ was included.
The model was simulated numerically, feeding it the
same coherence modulations that were used in the
psychophysical experiments and using the same proce-
dures for analysis. Model parameters were roughly
adjusted by hand to fit the results qualitatively. No
attempt was made to quantitatively match the data or to
build a model that explains a large body of other
experimental findings. Figure 5 shows the rivalrous
model behavior. Mutual inhibition in combination with
adaptation results in an unstable flip-flop in which
dominance regularly flips from one eye to the other. Note
that we did not introduce an internal noise source either,
which resulted in a quite narrow spread of interval
Yet, this simple model reproduces many of the
characteristics found in the real data (see Figure 6).
Profiles are biphasic and show a similar asymmetry for
modulations in the suppressed and dominant stimuli.
The timing of correlations obviously deviates from the
psychophysical data: the peak correlation occurs at t ¼ 0
because no response latency was introduced. Thus,
the difference in latency between model and real data
reveals the response latency between the occurrence
of internal (low-level) flips and the time of button presses.
Peak latencies for different observers varied from 380
to 620 ms (mean 560 T 127 ms, n ¼ 5). The zero
crossings for the nonsuppressed signals were very
consistent across observers (between 1600 and 1700 ms
before button presses). Zero crossings for the suppressed
signals showed a similar variation in latency as the
positive modulation peaks: varying between 900 and
We have shown that dynamic interactions between
rivalrous stimuli can be studied using a reverse correlation
technique. Flip-triggered noise averages reveal the bal-
ance between adaptation and mutual inhibition and its
dynamic change during the rivalry cycle. The important
conclusion is that biphasic behavior has to be taken into
account in perceptual rivalry studies in which signal
strengths are varied. Changing input signals at different
moments may result in opposite effects.
The biphasic behavior is not an artifact of the reverse
correlation procedure. Although the time average of the
noise trace by definition equals the mean coherence value,
only part of the noise, determined by the time of flips, is
analyzed. Most importantly, each signal is based on only
half of the total time, as determined by the dominance
flips. The flip-triggered averages thus do not necessarily
sum to the mean noise value. A comparison of model
results and psychophysical results shows that the slow
decay of the traces after the peak of the correlation
function must be due to the temporal correlation in the
noise signals. The model peaks sharply at time t ¼ 0, after
which modulations cannot affect the flip anymore. Thus,
decay after the peak must be due to temporal correlations.
This correlation is due to low-pass filtering of the external
noise. It should be noted that this slow decay might also
contribute to modulations in the earlier part of the
correlation functions. For the shorter interval durations,
these deflections fall within the time window over which
signals for the next flip are being averaged. Further model
Figure 5. Example of output signals for left and right eye.
Figure 6. Simulation results. Flip-triggered averages of coherence
values in the dominant eye (black curve) and nondominant eye
(green curve). The simulation results are shown in the same
format as the data in Figure 3.
Journal of Vision (2006) 6, 304–310Lankheet308
simulations, in which we skipped the first 1.5 s after flips
in the analysis, however, showed similar results. For the
experimental data, the effect is even smaller because the
slow decay is almost completely finished at the time of
button presses. Omitting 1.5 s after button presses from
the analysis slightly increased the variance for long
latencies but otherwise gave indistinguishable results.
We used independent modulations of signal strength for
the two eyes, which allowed us to study the asymmetries
in this process as well. Our finding therefore sheds new
light on the asymmetries in rivalry processes. Levelt’s
(1965, 1967) pioneering work clearly showed asym-
metrical contrast effects in the dominant and in the
suppressed eye. This finding has been a cornerstone for nu-
merous psychophysical and modeling studies (Matsuoka,
1984; Mueller, 1990; Stollenwerk & Bode, 2003; Wilson,
2003). The biphasic profiles that we measured suggest
that variations of fixed, continuous levels may largely
average out the underlying dynamic effects. Asymmetries
are not a static property; instead, our data show how the
asymmetry arises from opposite effects during different
phases of the rivalry cycle.
Variations on the method as presented here open many
new opportunities for studying dynamics and asymmetries
in both monocular and binocular rivalry processes.
Similar experiments could, for example, be performed
for luminance contrast variations of rivalrous gratings in
the two eyes. One can thus study and compare the
dynamics of different rivalry processes directly, rather
than indirectly comparing interval distributions. The
method also provides a precise estimate of the time
between the Binternal reversal[ and the moment of button
presses. Differences in the time required for different
rivalry mechanisms to reach visual awareness might thus
The reverse correlation method allows separating
dynamic effects resulting from dominant and suppressed
signals. This opens the opportunity to link, for example,
cognitive manipulation of rivalry in much more detail to
changes in the underlying mechanisms. Reverse correla-
tion in combination with a cognitive manipulation task
may very well reveal how observers are able to adjust
sensitivity for the different stimuli.
Observers clearly differed in their sensitivity to external
noise. RB very much behaves like the noise free model.
He shows a large amplitude effect, with relatively narrow
confidence intervals. MLO on the other hand shows much
smaller modulation amplitudes. The fact that RB has a
long mean interval duration and MLO a very short mean
interval suggests that this variation might correspond to
different levels of internal noise. It would also imply that
one could study the internal noise source by varying the
properties of the external noise. By measuring noise
thresholds for different temporal properties of the external
noise, we might be able to measure the temporal
frequency content of the internal noise.
We conclude that one should take the dynamics of
adaptation and inhibition into account in linking stim-
ulus manipulations and noise sources to interval distri-
butions. Reverse correlation provides an extra handle to
unravel the different processes involved in perceptual
I am grateful to Andre Noest for inspiring discussions.
Commercial relationships: none.
Corresponding author: Martin J. M. Lankheet.
Address: Padualaan 8, 3584 CH Utrecht, The Netherlands.
Alais, D., & Blake, R. (1998). Interactions between global
motion and local binocular rivalry. Vision Research,
38(5), 637Y644. [PubMed]
Blake, R., & Logothetis, N. K. (2002). Visual competi-
tion. Nature Reviews. Neuroscience, 3(1), 13Y21.
Blake, R., Yu, K., Lokey, M., & Norman, H. (1998).
Binocular rivalry and motion perception. Journal
of Cognitive Neuroscience, 10(1), 46Y60. [PubMed]
Britten, K. H., Shadlen, M. N., Newsome, W. T., &
Movshon, J. A. (1992). The analysis of visual
motion: A comparison of neuronal and psychophys-
ical performance. Journal of Neuroscience, 12(12),
4745Y4765. [PubMed] [Article]
Crick, F., & Koch, C. (1998). Consciousness and neuro-
science. Cerebral Cortex, 8, 97Y107. [PubMed]
Laing, C. R., & Chow, C. C. (2002). A spiking neuron
model for binocular rivalry. Journal of Computa-
tional Neuroscience, 12(1), 39Y53. [PubMed]
Lee, S. H., & Blake, R. (2002). V1 activity is reduced
during binocular rivalry. Journal of Vision, 2(9),
doi:10.1167/2.9.4. [PubMed] [Article]
Lehky, S. R. (1988). An astable multivibrator model of
binocular rivalry. Perception, 17(2), 215Y228.
Leopold, D. A., & Logothetis, N. K. (1996). Activity
changes in early visual cortex reflect monkeys’
Journal of Vision (2006) 6, 304–310Lankheet 309
percepts during binocular rivalry. Nature, 379(6565),
Levelt, W. (1965). On binocular rivalry. Institute for
Perception RVO-TNO: Soesterberg, The Nether-
Levelt, W. J. (1967). Note on the distribution of
dominance times in binocular rivalry. British Journal
of Psychology, 58(1), 143Y145. [PubMed]
Logothetis, N. K. (1998). Single units and conscious
vision. Philosophical Transactions of the Royal
Society of London. Series B, Biological Sciences,
353(1377), 1801Y1818. [PubMed]
Logothetis, N. K., & Schall, J. D. (1989). Neuronal
correlates of subjective visual perception. Science,
245(4919), 761Y763. [PubMed]
Matsuoka, K. (1984). The dynamic model of binocular
rivalry. Biological Cybernetics, 49(3), 201Y208.
Meng, X., Chen, Y., & Qian, N. (2004). Both monocular
and binocular signals contribute to motion rivalry.
Vision Research, 44(1), 45Y55. [PubMed]
Mueller, T. J. (1990). A physiological model of binocular
rivalry. Visual Neuroscience, 4(1), 63Y73. [PubMed]
Norman, H. F., Norman, J. F., & Bilotta, J. (2000). The
temporal course of suppression during binocular
rivalry. Perception, 29(7), 831Y841. [PubMed]
Polonsky, A., Blake, R., Braun, J., & Heeger, D. J. (2000).
Neuronal activity in human primary visual cortex
correlates with perception during binocular rivalry.
Nature Neuroscience, 3(11), 1153Y1159. [PubMed]
Stollenwerk, L., & Bode, M. (2003). Lateral neural model
of binocular rivalry. Neural Computation, 15(12),
Tong, F., Nakayama, K., Vaughan, J. T., & Kanwisher, N.
(1998). Binocular rivalry and visual awareness in
human extrastriate cortex. Neuron, 21(4), 753Y759.
van de Grind, W. A., van Hof, P., van der Smagt, M. J., &
Verstraten, F. A. (2001). Slow and fast visual motion
channels have independent binocular-rivalry stages.
Proceedings of the Royal Society of London.
Series B, Biological Sciences, 268(1465), 437Y443.
Wheatstone, C. (1838). On some remarkable, and hitherto
unobserved, phenomena of binocular vision. Philo-
sophical Transactions of the Royal Society of Lon-
don, 128, 371Y394.
Wilson, H. R. (2003). Computational evidence for a
rivalry hierarchy in vision. Proceedings of the
National Academy of Science of the United States of
America, 100(24), 14499Y14503. [PubMed] [Article]
Journal of Vision (2006) 6, 304–310Lankheet 310