Neuronal firing rate, inter-neuron correlation and synchrony in area MT are correlated with directional choices during stimulus and reward expectation.

Abstract

Sensation, memories, and predictions contribute to choices in everyday life, and their relative impact should change with task constraints. To investigate how the impact from sensory cortex on decision making varies with task constraints we trained macaque monkeys in a direction discrimination task where they could maximize reward by waiting for sensory visual information early in a trial, while focusing on memory and reward prediction as a trial progressed. The task constraints caused animals to indicate decisions in complete absence of visual motion stimuli (stimulus independent decisions), as 25% of the trials were 'no stimulus' trials. On 'no stimulus' trials reward delivery depended on the current decision in relation to the decision history. Stimulus independent decisions occurred during an epoch when a stimulus could in principle have been presented, or afterwards when stimuli could not occur anymore. Stimulus independent decisions were significantly different during these two periods. Reward exploitation was more efficient late in the trial, but it was not associated with systematic activity changes in directionally selective neurons in area MT. Conversely, systematic changes of neuronal activity and firing rate correlation in directionally selective middle temporal area (MT) neurons were restricted to a short time period before early decisions. Changing task constraints in the course of a single trial thus determines how neurons in sensory areas contribute to decision making.

Full text

Page 1

RESEARCH ARTICLE
Neuronal firing rate, inter-neuron correlation and synchrony
in area MT are correlated with directional choices during
stimulus and reward expectation
A. Thiele Æ Æ K.-P. Hoffmann
Received: 10 December 2007/Accepted: 13 April 2008/Published online: 29 April 2008
? Springer-Verlag 2008
Abstract
ute to choices in everyday life, and their relative impact
should change with task constraints. To investigate how the
impact from sensory cortex on decision making varies with
task constraints we trained macaque monkeys in a direction
discrimination task where they could maximize reward by
waiting for sensory visual information early in a trial, while
focusing on memory and reward prediction as a trial pro-
gressed. The task constraints caused animals to indicate
decisions in complete absence of visual motion stimuli
(stimulus independent decisions), as 25% of the trials were
‘no stimulus’ trials. On ‘no stimulus’ trials reward delivery
depended on the current decision in relation to the decision
history. Stimulus independent decisions occurred during an
epoch when a stimulus could in principle have been pre-
sented, or afterwards when stimuli could not occur
anymore. Stimulus independent decisions were signifi-
cantlydifferentduringthese
exploitation was more efficient late in the trial, but it was
not associated with systematic activity changes in direc-
tionally selective neurons in area MT. Conversely,
systematic changes of neuronal activity and firing rate
correlation in directionally selective middle temporal area
(MT) neurons were restricted to a short time period before
Sensation, memories, and predictions contrib-
twoperiods.Reward
early decisions. Changing task constraints in the course of a
single trial thus determines how neurons in sensory areas
contribute to decision making.
Keywords
Direction discrimination ? Middle temporal ?
Motion processing
Choice probability ?
Introduction
The brain uses various sources of information to form
decisions and guide behaviour appropriately. In order to do
so it needs to gate the flow of information in relation to the
subjects’ current goals. Under normal conditions sources of
information that affect information processing and decision
making are sensory cues (Britten et al. 1992), memory
traces (Miller et al. 1993, 1996; Rainer et al. 1998), abstract
response rules (Konishi et al. 1998; Wallis et al. 2001;
Nakahara et al. 2002; Wallis and Miller 2003; Mansouri
et al. 2006), attention (Desimone and Duncan 1995; Treue
and Maunsell 1996; Roelfsema et al. 1998; Roberts et al.
2007) or predictions (Sugrue et al. 2004; Yang and Shadlen
2007), and neuronal signatures of these have been descri-
bed in a variety of different areas and task constraints. An
area that has been extensively studied in terms of its con-
tribution to sensory based decision making is area MT of
the macaque monkey (Newsome et al. 1990; Salzman et al.
1990; Britten et al. 1992, 1996; Thiele and Hoffmann 1996;
Bradley et al. 1998; Thiele et al. 1999, 2001; Dodd et al.
2001). Neuronal activity in MT correlates with the choice
when animals expect visual motion input, even when the
input is not informative (Britten et al. 1996; Bradley et al.
1998), or entirely absent (Thiele and Hoffmann 1996). In
the absence of informative cues small activity changes in
Electronic supplementary material
article (doi:10.1007/s00221-008-1391-z) contains supplementary
material, which is available to authorized users.
The online version of this
A. Thiele ? K.-P.Hoffmann
Allgemeine Zoologie & Neurobiologie,
Ruhr-Universita ¨t Bochum, Bochum, Germany
A. Thiele (&)
Institute of Neuroscience, Henry Wellcome Building,
Newcastle University, Newcastle upon Tyne NE2 4HH, UK
e-mail: alex.thiele@ncl.ac.uk
123
Exp Brain Res (2008) 188:559–577
DOI 10.1007/s00221-008-1391-z

Page 2

area MT might be interpreted by downstream neurons as if
they were associated with a specific direction of motion,
thereby influencing decision making when monkey engage
in a direction discrimination task. Electrical microstimu-
lation experiments have shown that the influence of MT
activity on downstream areas is most effective during
specific epochs of a task; namely during epochs where
decision relevant information was present in the external
world, while it was less effective during externally cued
task epochs where such information was absent (Seide-
mann et al. 1998). This was interpreted as evidence for
selective gating of sensory signals in the cortex. It remains
to be determined how the flow of activity is gated under
more natural circumstances where changes of task con-
straints are gradual, not signaled by external cues (e.g.
stimulus on and offset), and where activity in MT was not
subject to artificial manipulation, but generated internally.
We were additionally interested in the mechanisms by
which decisions are formed. It is well established that
systematic changes of firing rates in MT contribute to
decision making in the absence of informative cues (for
review see Krug 2004). A complimentary mechanisms
which could benefit decision making is synchrony between
neurons that are tuned for similar characteristics, as inte-
gration of excitatory post synaptic potentials (EPSPs) at
downstream stages may benefit from synchronized inputs
(Fries et al. 2001). Although a previous report failed to find
a systematic relation between neuronal synchrony in MT
and the animal’s choice (Bair et al. 2001), a significant
correlation between the speed of behavioural reports and
gamma band synchronization in V4 during attentional
selection has been demonstrated (Womelsdorf et al. 2006).
If correlated activity contributes to decision making by
increasing the impact of inputs in downstream neurons, it
might be restricted to time periods where such inputs are
relevant for the animal (i.e. when they wait for a visual
stimulus), and might be reduced when directional choices
cannot benefit from sensory information.
To investigate how firing rate and neuronal synchrony in
area MT contribute to directional choice under changing
task constraints, we trained monkeys to participate in dis-
criminating the direction of moving stimuli or indicate
decisions once the probability of stimulus occurrence had
returned to zero for that particular trial. For the latter the
probability of rewards depended on the animal’s choice
history. They thus had to change strategy and use memory
traces if they wanted to maximize reward income. We
found that firing rate, firing rate correlation and synchrony
between simultaneously recorded neurons significantly
correlated with the choice direction while animals waited
for sensory stimuli. When an animal’s decision was more
based on an attempt to maximize the reward income, MT
activity, firing rate correlation and synchrony were not
systematically related to the choice. Thus, the presence or
absence of choice related activity pattern in area MT varies
with task constraints within the time course of single trials.
Methods
All experiments were carried out in accordance with the
European Communities Council Directive 1986 (86 909
EEC) and National Institutes of Health guidelines for care
and use of animals for experimental procedures. Details
regarding training, surgery, and post-operative care are
given in (Thiele et al. 1999). Two male Macaca Mulatta
participated in the study. Data from one of the two mon-
keys (AR) have been published previously in a different
context, where error trials in relation to stimulus induced
neuronal activity were investigated (Thiele et al. 1999). No
aspects of any of the data from the second monkey (CS)
have been published previously.
Stimuli and behavioural paradigm and controls
The monkeys were trained in a direction discrimination
task (Fig. 1a). During the experiments, each animal was
seated comfortably in a primate chair with its head
restrained. We monitored eye movements using scleral
search coils (optimum resolution of 1.50), sampled at
500 Hz. Monkeys started a trial by clutching a central
touch bar in front of its chest, upon which a fixation point
(0.2? in diameter) was back projected by a light emitting
diode onto a translucent tangent screen. The screen sub-
tended 90? 9 90? of visual angle at a viewing distance of
38 cm. The fixation point was always presented in the
centre of the projection screen. The maximum fixation
window was ±1?. Monkeys were required to fixate within
500 ms after the appearance of the fixation point.
Stimuli
During each experiment, the time of stimulus onset, con-
trast, and direction were varied randomly. The direction of
motion was along one of the four cardinal directions.
Stimuli were presented in the receptive field of the recor-
ded neuron. When two or more single units were recorded
simultaneously, the stimulus covered all of the units’
receptive fields. Stimuli consisted of square wave gratings
(0.3–0.5 cyc/deg) moving uni-directionally within a qua-
dratic aperture. During the whole experiment, a stationary
gaussian-filtered white noise stimulus was also back pro-
jected onto the screen (by a slide projector). This added
stationary noise made the task more difficult, causing a
higher percentage of error trials (Hoffmann and von Seelen
1980). The noisy stationary background was generated by
560 Exp Brain Res (2008) 188:559–577
123

Page 3

putting an on-axis Fourier hologram of a piece of ground
glass with Gaussian amplitude density distribution and
constant power between 0 and 7 cyc/deg (von Seelen and
Hoffmann 1976) into the beam of the slide projector. The
mean intensity of the noise background was 1.7 cd/m2with
a RMS contrast of 0.3053.
The stimulus contrast was varied from high contrast
levels to invisible contrast (0%). Four contrast levels were
tested for each neuron (17, 4, 2, and 0% contrast). Stimuli
were back-projected with an EPS 4000 video projector
(Electrohome) onto the translucent tangent screen, super-
imposed onto the stationary noise background. The
stimulus and background intensity were measured using a
photo-multiplier (EMI,14 dinodes, 20S, aperture 0.04? of
visual angle), and the linearity of measurements was
ensured with 50% transmission neutral grey filters (Schott,
Mainz).
Behavioural paradigm
The monkeys performed a reaction time task. As soon as
they perceived (or believed they perceived) the direction of
motion, they had to release the central bar and touch one of
the four peripheral bars. These were positioned according
to the directions of motion (see Fig. 1a). The release of the
central touch bar was taken as the reaction time (RT). A
‘go-signal’ was never presented. After touching the
peripheral touch bar, the monkey had to keep fixation for
another 500 ms, during which the stimulus continued to
move through the receptive field (provided it was
displayed).
Stimuli could move up, right, down, or leftwards within
the aperture. Monkeys had to discriminate the direction of
motion and indicate it by a corresponding hand movement
to one of four touch-bars surrounding a central touch bar
located in front of the animals. The animals could respond
immediately when they saw a stimulus, i.e. they engaged in
a reaction time task. Once a stimulus was presented they
had 2,500 ms to indicate the direction of stimulus motion.
On ‘no stimulus’ trials they had 5,000 ms to indicate a
decision. Thereafter the trial was aborted without reward
and the animal could start a new trial. On trials where they
indicated the correct direction of stimulus motion they
received a fixed quantity of 0.1 ml of juice for correct
motion discrimination. Stimuli could occur 640, 1,240,
1,840, 2,440, and 3,040 ms after trial onset for monkey
AR. Two different time series were used for monkey CS:
(a) the stimulus could occur after 520, 1,020, 1,520,
2,020 ms, or (b) in a second set of experiments it could
occur after 1,020, 1,520, 2,020, or 2,520 ms.
The probability of stimulus presentation, the stimulus
luminance contrast, direction of motion, and its presenta-
tion time within a trial were all entirely randomized on a
trial by trial basis, thus making the visibility of the stim-
ulus, and its direction of motion entirely unpredictable,
while its time of occurrence was only predictable in
accordance with the hazard function (see below). The
uncertainty of whether and when a stimulus would be
0
5000
0
20
40
60
80
100
0
5000
0
1
0
5000
5
15
25
35
0
5000
05000
20
40
60
05000
0
1
0
1
frequency
cumulative density
time after trial onset [ms]
A
B
monkey AR
monkey CS monkey CS
Fig. 1 a Experimental task. Monkeys faced a transparent screen onto
which a stationary noise background was back-projected by a slide
projector and onto which the stimulus was back-projected by a video
projector. When visible stimuli (bars to the lower right of the
monkey’s fixation) were presented monkeys had to indicate the
direction of motion with a hand movement to a corresponding touch
bar. Stimulus motion direction, probability of stimulus onset time and
visibility were randomized such that high levels of uncertainty about
stimulus occurrence caused animal to indicate stimulus independent
decisions (SIDs, ‘‘Methods’’). In the current paper all our analyses
exclusively deal with SID trials, i.e. when no visual stimulus was
presented. b Distribution of SID timing (black solid curve, upper
row). SID timing for the four different directional choices are shown
separately below (grey and dotted lines). SID times closely coincided
(or followed) the possible stimulus presentation times, as evident by
the small peaks in the distribution of SID timings. The step function
shows the cumulative hazard function (normalized to 1). The lower
row shows the cumulative hazard function (dashed line), the
cumulative hazard function convolved with a Gaussian with 150 ms
sigma (smooth line on top of the cumulative hazards function) and the
normalized cumulative SIDs as a function of time after trial onset. In
both animals the cumulative SID functions closely trailed the
cumulative hazard function
Exp Brain Res (2008) 188:559–577561
123

Page 4

presented ensured that animals often indicated directional
decisions in total absence of visual motion. Since these
decisions occurred in complete absence of a moving visual
stimulus we termed them stimulus independent decisions
(SIDs). SIDs in the context of this paper could occur during
‘no stimulus trials’ as assigned by the computer. Addi-
tionally, SID trials within the context of this report were
also trials where a stimulus should have been presented, but
animals indicated a decision before the intended stimulus
onset. If SIDs occurred before an intended stimulus pre-
sentation, the stimulus presentation was abandoned, i.e. the
animals also indicate a decision in complete absence of a
visual motion stimulus. These SIDs were caused by an
‘early’ (or anticipatory) choice of the animal, not by the
computer generated trial selection. They were never
rewarded, but are included in the paper as they are identical
in terms of the physically display on the screen (the sta-
tionary structured background), and thus in terms of retinal
stimulation. SIDs that occurred during computer generated
‘no stimulus’ trials were rewarded on 50% of the trials on
average, but this probability of reward and the reward
magnitude was adjusted according to the animal’s choice
history (see below). This paper exclusively focuses on
trials where SIDs of either type occurred.
Adjustment of reward probability and magnitude
Reward probability and magnitude were only adjusted for
SIDs during computer generated ‘no stimulus’ trials, not
on animal generated SID trials. The latter were never
rewarded. Reward probability on computer generated ‘no
stimulus’ SID trials was adjusted such that choice bias was
reduced. At the same time it ensured that the animals
could maximize their reward income on these SID trials
by employing a strategy that kept track of previous choi-
ces and the associated income. Provided the animals learn
these contingencies it would allow to determine whether
directional choices are reflected differently in neuronal
activity when choices are made while animals wait for a
sensory stimulus, compared to when the identical choice is
made based on the reward history. On trials when the
computer selected a ‘no stimulus’ condition animals were
randomly rewarded on 50% of the trials (mean) for SIDs.
Reward probability on computer generated SID trials was
adjusted according to the animal’s past SIDs, i.e. the
animal’s decision history. The decision history consisted
of the last 1,000 SIDs that had occurred, irrespective of
whether it was a SID in relation to a no stimulus trial, or a
SID that preceded an intended stimulus presentation. On
SID occurrence the decision history was renewed in a ‘first
in–first out’ manner, i.e. the SID 1,000 trials ago was
eliminated from the buffer, the SID in position 999 was
moved to position 1,000, position 998 was moved to
position 999, etc.. The SID history was stored to the
computer disk, such that on each day the last 1,000 SIDs
that had occurred during the previous session(s) were
initially uploaded. SIDs that had occurred more recently
were given more weight in calculating the current reward
probability. The last 10 SIDs (block 1) had a weight of
0.5, SIDs 10–100 (block 2) had a weight of 0.3, and SIDs
100–1000 (block 3) had a weight of 0.2. The exact
probability of a reward for a certain SID was calculated
according to:
?
where p(R)B is the reward probability for a certain SID
given the current block (1–3), dp is the number of SIDs in
the current block matching the current SID direction, sB is
the size of the current block (10, 90, 900), and nd is the
number of different SIDs possible (i.e. 4). Given that the
different blocks of past decisions had different weights, the
total reward probability was calculated according to:
pðRÞB¼
0:5 þ 0:5 1 ? dp=ðsB=ndÞ
0:5 ðsB? dpÞ=ðsB? sB=ndÞ
?
??jdp?sB=nd
jdp[sB=nd
?
pðRtotalÞ ¼ pðRÞB10:5 þ pðRÞB20:3 þ pðRÞB30:2
where the three different p(R)s are the reward probabilities
based on the three blocks multiplied by the weighting
factor. In addition to adjusting the probability of a reward,
we also adjusted the reward magnitude based on the
monkeys’ SID history. This ensured that animals could in
principle estimate the current reward probability of the
choice from the reward magnitude (a large reward is
equivalent to high probability), and thus adjust choice
accordingly (i.e. resample if a large reward was given,
avoid this choice for a while if a small reward was given).
Reward magnitude was adjusted by changing the solenoid
opening time, i.e. adjusting the duration of reward delivery.
In addition to linear increases/decreases of reward
magnitude depending on choice history a step non-
linearity was introduced for SIDs occurring at the
transition of \50% reward probability and C50% reward
probability. The solenoid opening time was calculated
according to
?
where OT(p) is the opening time as a function of the cur-
rent reward probability (p(Rtotal)) given the animals’
decision. The amount of reward was linearly related to the
opening time from 50 ms opening time onwards, whereby
a 50 ms opening time yielded 0.05 ml per reward, and
every 1 ms increase added 0.001 ml to this. Opening times
below 50 ms resulted in non linear decreases (accelerating
decreases), such that at 25 ms opening time a reward
yielded 0.01 ml, and basically no reward was delivered at
opening times of\10 ms due to the inertia of the solenoid.
OTðpÞ ¼
100 þ 100pðRtotalÞ
100 ? 100 0:5 ? pðRtotalÞ
??jpðRtotalÞ?0:5
jpðRtotalÞ\0:5
562 Exp Brain Res (2008) 188:559–577
123

Page 5

SID times and hazard function
For each monkey the SID times were determined as the
time points when the hand disconnected from the central
touch-bar. Thus, we refer to the decision time as the hand
movement onset, although the decision itself will have
occurred somewhat before the movement onset. The fre-
quency of decisions as a function of the direction of the
SID and as a function of time from trial onset was calcu-
lated. Additionally we determined the stimulus occurrence
probability (the probability density function of stimulus
occurrence), by calculating the frequency distribution of
stimulus presentation as a function of time, and divided this
distribution by the total number of trials (i.e. those when a
stimulus was presented and those when no stimulus was
presented. The relevant probability for the animal, of
whether to indicate a decision if no stimulus has yet been
detected is represented by the hazard function (h(t)). This is
the probability that a stimulus will occur at any time given
that no stimulus has been presented yet:
hðtÞ ¼
pðtÞ
1 ?RpðtÞdt
where p(t) is the probability density function of stimulus
occurrence, and the integral corresponds to its cumulative
counterpart, the cumulative density function of stimulus
occurrence (Smith 1995).
Behavioural controls
Throughout the trial (and until 500 ms after the indication
of a decision) animals had to fixate the fixation spot within
a window of 2? side length, i.e. ±1? from the fixation spot.
Eye position was monitored by means of scleral search
coils (optimum resolution of 1.50), sampled at 500 Hz.
Electrophysiology and assessment of basic neuronal
properties
Recordings were performed using the ‘Eckhorn Matrix’ (7
channel, Uwe Thomas Recording, Marburg, Germany).
Glass-insulated platinum–iridium electrodes (Uwe Thomas
Recording, Marburg, impedance: 1.5–3 MX at 1 kHz)
wereadvancedthroughguidetubes
305 lm) into the brain. Up to seven guidetubes were
inserted per recording session, the guide tubes were
arranged in a circular array with one guidetube in the
centre of six surrounding tubes (overall diameter was
915 9 915 lm). Each guidetube was sharpened, such that
all inserted guidetubes together formed a single tip.
Amplified electrical activity from the cortex was band-pass
filtered (0.3–10 kHz), and passed through oscilloscopes to
Schmidt trigger thresholding (built in house) and spike
(outerdiameter
sorting devices (Alpha Omega, Nazareth, Israel). In mon-
key AR we focused on single unit recording (all cells
reported here were subject to spike sorting), while in
monkey CS we focused on single unit and multi unit
recordings, whereby signals from some electrodes were
subjected to spike sorting, while others were intentionally
subjected to Schmitt-trigger thresholding, which resulted in
multi-unit signals. The latter was done as synchrony
between recording sites is more easily detected from multi-
than single-unit recordings (Kreiter and Singer 1996; Ro-
elfsema et al. 2004). For single unit recordings the quality
of spike separation was controlled online by displaying the
interspike interval distribution for each spike channel on
the monitor of the recording personal computer.
Prior to recording, the receptive field location of each
cell was mapped using a hand-held projector while the
monkey fixated a central target on a dark background. Data
were recorded only after the monkey had become well
adapted to the background luminance (*0.5 cd/m2) for
approximately 20–30 min. In addition, the cell’s spikes
(single and multi- unit activity) had to be well isolated
(stable) for at least 5 min while the monkey performed the
task before data acquisition began.
Direction selectivity
Initially we determined for each neuron whether it was
directionally selective by calculating the direction index
(DI). This was based on the neuronal activity following the
period of 50–300 ms after a 17% contrast stimulus had
been presented and the animal made a correct decision. For
the calculation of the direction index, the direction of
motion that yielded the largest activity was termed the
preferred direction (PD), the direction of motion opposite
to this direction was termed the anti-preferred direction
(APD). The direction index (DI) was calculated according
to:
DI ¼ 1 ?actAPD
actPD
where actAPD was the neuronal activity associated with
anti-preferred direction of motion, actPD the neuronal
activity associated with preferred direction of motion after
subtraction of background activity. If the DI exceeded 0.5,
the neuron was taken as direction selective. Only direction
selective neurons are included in the current paper. For
direction selective cells we then calculated the ‘vector
average’ preferred direction. This preferred direction was
the vector average across the four directions of stimulus
motion at 17% luminance contrast as described previously
(Thiele and Hoffmann 1996). It was calculated to analyse
whether cells with more similar preferred direction showed
stronger synchrony and rate correlation (see below). Due to
Exp Brain Res (2008) 188:559–577563
123